Artificial intelligence-aided design: Smart Design for sustainable city development

Abstract

Current planning and design decision support systems show limitations in the integration of design, science, and computation. Planning support systems with manual design and post-design evaluations impose major challenges in exploring huge design spaces. Generative design systems largely neglect the wicked nature of design problems and lack appropriate representation methods and simulation tools at the urban scale. To tackle those challenges, this research developed a Smart Design framework featuring urban design decision-making reinforced by artificial intelligence-aided design (AIAD). The Smart Design framework treats urban design as an emergent pattern formation processes with contextualized and dynamic objectives. The framework integrates design thinking, advanced artificial intelligence search techniques (e.g. genetic algorithms), urban scale performance simulations, and participation to better inform decision-making. Through four major stages, the framework combines the ideas of Science for Design and Design in Science. The significance and potential of the Smart Design framework are demonstrated in an urban design study of Gangnam superblocks in Seoul, South Korea. The study explores sustainable urban forms in the high-density, super-complex, and hyper-consumptive environment of Gangnam, which can also be found in many other Asian contexts. The case study illustrates how the framework identifies design solutions for sustainable city development in the process of participatory decision-making through the co-evolution of design problems and solutions.

Keywords

Smart Design artificial intelligence-aided design (AIAD)genetic algorithms energy efficient urban form planning support systems generative design

Introduction

Increasing computational power and advancing computational technologies, such as artificial intelligence (AI), provide new opportunities to tackle city issues (Batty, 2018), especially the sustainable development challenges in environmental protection, economic growth, and social justice widely recognized for a long time (Campbell, 1996). While many scholars have been exploiting these computational advantages to study cities, most of their focuses were on the scientific findings that can be used to support planning and design as Science for Design (Batty, 2013a). But sustainable urban development also requires discussions on how design itself should proceed as Design in Science, which is less well understood (Batty, 2013b; Lynch, 1981; Simon, 1969). An emerging question is: how can we plan and design sustainable cities with a smarter design method that combines both Science for Design and Design in Science?

The urban design process, however, is often discussed with reference to two distinct meanings in the literature under the same terminology. The first being the general design about the overall design process from problem definitions to solutions, as the big-D “Design” (Goodchild, 2010). The second refers to the narrow design that focuses on a well-defined design task to generate solutions usually as part of the general design process, as the small-D “design” (Goodchild, 2010). Both processes have their characteristics and challenges. On the one hand, the general design process deals with a series of activities that are ill-structured (Simon, 1969) and wicked as in planning (Rittel and Webber, 1973). A common dilemma is that the design process toward solutions is dependent on problem definition, which often gets clarified during the design process (Buchanan, 1992; Rowe, 1987; Simon, 1969). Participation from different stakeholders adds another layer of complexity (Batty, 2013b). The narrow design process, on the other hand, can be understood as searching in a well-defined design space for a particular design problem. How to search more efficiently and effectively is the main focus of this process.

Two schools dominate the prevailing planning and design decision support systems that facilitate the design process. The planning support systems (PSS) school, which is defined as loosely coupled technique assemblages to aid planning decision-making (Batty, 2003), and the generative design systems school that produce novel and efficient spatial designs through computing (Shea et al., 2005). Both of them have advantages and limitations in supporting complex urban sustainable design issues with design, science, and computation (Goodchild, 2010; Hopkins, 1999). Especially, their focus on the two types of design processes and the application of scientific tools for these processes are quite different.

The school of PSS focuses on the general planning and design process and how various computational tools can be used to support components in the process (Klosterman, 1997). Following the evolution of the view of planning processes, the general process in PSS developed from system optimization to collective design (Klosterman, 1997), and from the linear and rational models to the recursive and communicative models (Batty, 2013b; Vonk et al., 2007). However, the narrow design process in PSS often adopts simple and limited methods, such as scenario making based on current trends and planners and designers’ expertise. Decisions are then made based on impact assessments of scenarios using scientific knowledge and tools (Geertman et al., 2015; Quan et al., 2013). Such post-design evaluation approaches have been argued to be limited in solving complex design issues in cities with nearly infinite design spaces because of large numbers of variables and rules, often resulting in suboptimal solutions (Rowe, 1987). The Geodesign method can be seen as the most recent model in this school, which “tightly couples the creation of design proposals with impact simulations influenced by geographic contexts, system thinking, and digital technology” (Steinitz, 2012). While Geodesign pushes the complexity of the general design process to a higher level, its treatment of the narrow design process largely follows the simplicity tradition in the PSS school. Under such a framework, the proposed and selected design solutions are usually suboptimal with manual design and the post-design evaluation often used in the impact model (online supplemental Figure 1a).

The other school of generative design systems stemmed from architectural and engineering design. The focus is more on physical form representation and computational support for the narrow design process (Chase, 2005; Knight and Stiny, 2001; Oxman, 2006). The recent development of the performance-based generative design method allows a more systematic search in the design space with scientific performance simulations and optimizations algorithms (Janssen, 2006; Krish, 2011; Oxman, 2009). However, most studies in the generative design school treat the general design process the same as the narrow design process, which has a clearly defined design problem and rational process. The co-evolution of design problems and solutions and the participation are largely missing. A typical generative design model considers the design development process as generally linear and rational, and from only the designers’ perspective (see online supplemental Figure 1b). Spatial scale is another challenge for generative designs to be applied to urban areas. New solutions are required for urban form representation and urban scale performance simulations to deal with increasing complexity when the urban form scales up. Solutions for such complexity should address interactions between the spatial form, social activity, and infrastructure system (Graham and Marvin, 2002).

Computational advances, especially AI development that aims to understand, and further design and build cognitive systems (Goel and Davies, 2011), have been greatly enhancing the capability of planning and design support systems in both schools. However, the aforementioned fundamental shortcomings remain in the two schools. Planning and design with AI have been a prominent topic, and their history is almost as long as that of AI (Batty and Yeh, 1991), with a focus on conjecturing the process of planning and design, and constructing computational tools to aid the process (Gero, 1991). While some works on narrow design generation gained success, growing ambitions to automate the general design process without acknowledging its wicked nature led to controversies and unfavorable results in early works, and the idea of design with AI became more questioned (Colgan et al., 1991). In recent years, AI regained its popularity with advances in technologies, computational power, and data availability. AI-aided design (AIAD)—which emerged in generative design—has also gained popularity (Martins et al., 2014; Schüler et al., 2017). The new generation of AIAD applies goal-oriented optimization algorithms to urban design, but the structure is not changed, and hence, they inherit the same strengths and limitations of the generative design school.

In the new era of AI, how can we overcome the aforementioned shortcomings of the two schools and better integrate design, science, and computation to aid our planning and design for sustainable city development? What role can AI play in such a process? This paper aims to answer these questions by developing a new Smart Design framework with enhanced design decision-making reinforced by AIAD. The application of AI in such a framework focuses on the narrow design generation process, while the general design process is still coordinated by designers. In the current AI fever, there are increasing concerns and fears over universal AI that could supplant most human workers. Although AI will replace some of the jobs that rely on repeated works, studies on AI and the human cognition system suggest that many of those fears are exaggerated or even unfounded (Goel and Davies, 2011). Current computers rely on serial processors that are well suited for rational and serial processing. AI developed on computer systems can outperform humans on rational and specific tasks such as the narrow design generation, but does not perform as well in complex cognitive tasks such as communication that permeate the general design process (Goel and Davies, 2011). Moving toward universal AI that can be used for general design requires breakthroughs in neuroscience to illuminate the nature of those cognitive processes, which may take a very long time (Goel and Davies, 2011). While the general design process in Smart Design does not utilize AI directly, narrow design generation can benefit greatly from AI techniques.

Smart Design framework

The Smart Design framework has two closely linked parts of the processes and the participants (see Figure 1). The processes part contains two levels of design processes. The first level is the general design process where computational tools and participants work together to guide the co-evolution of design problems and solutions, and the second is the narrow design generation where computational tools and AI techniques such as search algorithms are used to support well-defined design problems. At the first level, the Smart Design framework takes the view of urban design as emergent pattern formation processes with evolving goals in synthesized spatial, social, environmental, and engineering systems. This general design process is organized into four major stages: human problem initialization, human–system interface, system optimization, and human–system interaction. The second level corresponds to the human–system interaction and system optimization stage. At this level, a particular design task is defined temporarily with the design objective and schema and to be solved by advanced AI search methods. Through the organization of the two levels of design, the Smart Design framework integrates system optimization and human knowledge and judgment to improve the efficiency and effectiveness of design practices, echoing with what Batty (2018) suggested for the next generation of AI development in planning and design.

Figure 1.

The Smart Design framework.

In the participants’ part, actual participants are determined by the involvement and collaboration in the design project. Generally, there are four distinct groups of experts involved, namely the public group including the citizens, the design group including designers, engineers, and computer scientists, the regulation group including government agents, and the implementation group including the developers and banks. They are empowered through the four stages to contribute to the design project with their knowledge and raise their concerns collaboratively, following the approach of urban living labs (Bulkeley et al., 2016; Niitamo et al., 2006). While narrow design generation requires only monitoring from designers and computer scientists, the general design process invites stakeholders’ participation through presentations and routine design meetings. Those design meetings are driven by the framework’s workflows and focused on the required components in the framework such as design goals and constraints. Each of those meetings is streamlined as two parts: in the first part, the public, the government, designers, engineers, etc. negotiate around the design problem definition or the feasibility of design solutions; the second part is featured by the inclusion of computer scientists in the meetings to help translate and encode the consensus results into quantitative elements as inputs for computing systems such as objective functions and parameters. Those meetings are organized and coordinated by planners and designers.

A visual illustration of this new Smart Design framework is given in Figure 1, featuring major components in the framework. The detailed models and participants in the four stages, as well as their interaction and loops, are discussed as follows.

Models and participants in four stages

Human problem initialization stage

The design begins with a dilemma: it should start with a well-defined problem but the problem is to be clarified through the exploration in the design process (Rowe, 1987; Simon, 1969). To break into this mutual reference cycle and begin the design, an initial interpretation of design problems is taken as the starting point. In the initial model in the first stage of the framework, human initial problem-solving, an interpretation of the initial problem and goal is proposed in the model based on a rough understanding of the pressing challenges in the design task. For example, the design of a downtown neighborhood may start from designers’ understanding of the most important problems, such as traffic congestion and lack of public space. The preliminary proposal is often not comprehensive but it can make the first dent in the problem-solving process and is expected to be modified and improved through the process. Human problem solving takes a major role in this model.

Human–system interface stage

In the second stage, the human–system interface, the descriptions of the design problems are encoded in mathematical representations within the objective model and the representational model. In the objective model, the visionary descriptions from the first stage are examined and further developed into objective functions, e.g. minimizing energy use. Also, constraints related to the objective functions are identified. The objective functions and constraints set criteria for design evaluations, which is critical and necessary in using advanced AI search methods to search for good designs. This model requires expert knowledge from designers, engineers, and computer scientists.

In the representational model, the actual design of the urban form is abstracted and represented as the basis for the design generations in the later stages. This representation follows a specific structure and conventions to match design problems and objective functions, and to embody the intended design schema that reflects a certain situation and language under the given design contexts (Rowe, 1987). Numerical parameterization is often used to model the representation of urban form to enable new design variations to emerge. Constraints in the parameterization are also identified as the range of parameter values and particular rules in their relationships. In this representational model, designers and computer scientists play a major role by identifying and coding the parameters together.

Both models serve to translate urban design goals, variables, and constraints into mathematical representations, though some of the translations are limited and subject to personal opinions and expertise of designers and computer scientists. Human knowledge such as domain knowledge and creativity are coded to support the advanced AI search techniques in later stages.

System optimization stage

The generative model and the optimization model are created in the third stage, system optimization. These models are the computational algorithms that drive design exploration. In the generative model, based on the parametric representation and constraints of urban form, different designs are generated given sets of parameters. The design variations are further parsed into geometric and semantic objects.

The optimization model uses advanced AI search techniques such as the genetic algorithms (GAs) (Russell and Norvig, 2016) and scientific performance simulation tools to guide the exploration of the design space toward optimal solutions. During the optimization process, the best fit individuals, or designs, are determined by their performance. And scientific simulation tools are used to evaluate those performances. The tools can incorporate a wide range of performance evaluation criteria for sustainable urban development, including economic, environmental, transportation, and visual quality performances.

In both the generative model and the optimization model, experts in design, engineering, and computation are greatly involved to integrate design generation, scientific modeling, and computational algorithms to search for optimal designs.

Human–system interaction stage

In the last stage, human–system interaction, the results from the system optimization stage are interpreted and visualized in 2D and 3D manners, or even in virtual and augmented environments for all participants to comprehend and interact for better decision-making, e.g. the CityScope platform developed by MIT Media Lab, which combines computational and physical models (Alonso et al., 2018).

The interaction between the results from system optimization and the discussions between stakeholders exhibits a network structure rather than the linear structure for only designers in Oxman’s generative design framework (Oxman, 2006). Through communication informed by visual and semantic representation, all four groups of stakeholders are able to decide which one among all candidate optimal designs is chosen, or whether new objectives and constraints need to be introduced to commence a new iteration from the second stage. The decision can be made through different types of formal consensus-reaching methods such as the classic and effective Delphi method.

Looping between four stages

This Smart Design framework is not simply a sequential model of the four stages. Instead, there are four internal loops. Two of them are recursions within the second and third stages—between the objective and representational models and the generative and optimization models, respectively. The other two loops are from the third and fourth stages back to the second stage, and these allow dynamic updates to the previous definitions of the design problems and representations at any given iteration during the design process. These four loops allow the design problems and solutions to co-evolve to support ill-structured design activities (Simon, 1969).

The framework has the following advantages over previous support systems. First, it emphasizes design thinking with an iterative and recursive structure to support the co-evolution of design problems and solutions. Second, it integrates the design generation, scientific evaluation, and advanced AI search techniques such as GAs to guide the exploration of the design space. Third, it provides a framework for system optimization and human knowledge to collaborate in the design process. Fourth, it invites the participation of stakeholders through formal stages during the entire design process and facilitates participation with the interfaces between the figurative design and abstract computing, which allows better-informed decision-making with computation results.

The framework is a container that could include various definitions of goals, constraints, parametric representations, methods, and participants to adapt to different situations. Therefore, with different detailed components, it can be applied to a wide range of urban planning and design tasks, from new town development to urban redevelopment.

Case study: Gangnam superblocks in Seoul

Site condition

Gangnam generally refers to the urban region that is south of the Han River in Seoul. The region was developed in the late 1960s when the population of Seoul began to rapidly grow along with the urbanization and industrialization of Korea at an unprecedented rate. This firmly determined yet underprepared rapid development was manifested in a group of superblocks, each approximately measuring 800 m × 800 m, much larger than the urban blocks commonly found in Western countries. Significantly, diverse urban conditions coexist within each superblock, and the unique morphology of Gangnam superblocks supports the diversity (Peponis et al., 2016). A figure-ground diagram of a typical superblock in Gangnam is given in Figure 2(b).

Figure 2.

(a1) to (a6) Structured representation of superblock and (b) a figure-ground diagram of a typical Gangnam superblock.

Gangnam superblocks were developed in three distinct stages: development of the residential buildings in the interior of the superblock, development of the commercial properties along the major internal streets, and development of the high-density offices on the periphery of the superblock (Yuh and Ahn, 2006). According to the current zoning regulations, the maximum allowable floor area ratio (FAR) in the commercial zones is typically 8, whereas in the residential zones is 3 (Seoul Metropolitan Government, 2018). Although there is no specific building height limitation, most skyscrapers in Gangnam are less than 30 stories.

Design problems and solutions: Co-evolution in two process iterations

Even though Gangnam is a major business center, its current FAR is well below the zoning limit. Given its high land prices, Gangnam’s future urban redevelopment will target intensification that will introduce many environmental and energy problems as has been happening in Seoul at large (Kim and Choe, 2011). Inclusion of environmental goals in development as in the Seoul 2030 Plan (Seoul Metropolitan Government, 2016) may lead to other challenges which require further modification of goals.

This case study simplifies the densification of superblocks as a hypothetical design for tabula rasa to demonstrate the framework. While the actual redevelopment will be more incremental and dynamic, a long-term picture can help to direct such a process. With additional information on parameter ranges and constraints identified, the framework is also able to support the design for incremental redevelopment. The case study is conducted in two iterations. In the first iteration, maximizing FAR is selected as the primary objective to meet the local demands on densification for better economic outcomes. The second iteration considers energy performance in its design problem definition as one of the environmental challenges brought by densification, with a deepened understanding of the design space from the first iteration.

The first iteration: Exploring the initial design problem

The first iteration explores the initial design problem by defining a starting design objective, developing an initial representation of a design schema, and determining design searching methods. The initial objective is defined as maximizing the FAR to its zoning limit. Because local zoning only specifies the net FAR of each parcel depending on the land use, the FAR limit for the entire superblock varies with the superblock design. This is again an example highlighting the wicked nature of design: the goal depends on the solutions. To tackle this dilemma, an initial representation was proposed to represent a stereotyped situation. The representation then underwent an updating because its resulting design variations included impractical patterns in the local context. With the updated representation, the optimization methods are applied and compared for this particular design problem. Found solutions led to discussions and reflections on the initial design problem.

Representation

The morphology of Gangnam superblocks is decomposed and understood in terms of various scales, namely superblock, region, block group, block, parcel, and building scales (Figure 2(a1) to (a6)). The formal principles are postulated from a brief survey of a series of historical aerial photographs of Gangnam taken since the early 1970s, and based on the general formal principles proposed by Peponis et al. (2016). The formal structure reflects the evolution of the city through both the top-down approach with building the regular and fast major street grid, and the bottom-up approach with the development of local streets following the historical and cultural urban fabrics, to adapt to modern transportation.

The design schema was structured in eight generative steps that reflect the development process (Figure 2): initialization, anchor-development, peripheral blocks, interior blocks, parcels, building footprints, building envelopes, and finalization. For each step, a set of parameters were assigned to specify a sequential instance (online supplemental Tables 1 and 2), and those instances cumulatively generate a Gangnam superblock design.

The first step operates at the “region” scale. It initiates a generative process by specifying the boundary, peripheral ring, and interior of a superblock. The peripheral ring is assumed to be commercial development, and the interior is assumed to be mainly a residential area. The second step focuses on the block group scale. It specifies an obliquely oriented anchor block group near the center, which contains the public center. Two traversing streets divide the superblock into 12 peripheral block groups, one anchor block group, and four interior quadrants. The third and fourth steps operate at the urban block scale. The third step specifies block arrangements within the peripheral block groups, and the fourth stage specifies independently oriented rectilinear grid arrangements within the interior quadrants. The fifth stage operates at the parcel scale. It specifies divisions of parcels into blocks, whose shapes could be convex triangles, squares, rectangles, or right trapezoids. Then each block can be divided into a 1 × 1, 1×n, or 2×n quad-grid configuration based on its local coordinate system and bounding rectangle. The sixth and seventh steps operate at the building scale. The sixth step specifies building footprints within parcels, with assigned conventions of setback requirements for simplicity. The seventh step specifies building heights, with simple assumptions of prism-like building envelopes and single building height for each block. The eighth step finalizes the design generation as both a 3D geometric model and a semantic model. The geometric model consists of vertices, edges, and faces, as well as properties such as land use, for all the geometries in the model, and the semantic model includes objects at various scales.

Such a design schema was then developed in the interface stage as a class with the object-oriented programing. Designs can be generated as objects of the class based on the values of 26 2D shape parameters and various height parameters.

Objective function

A superblock is typically planned as a whole development unit, and the FAR of internal parcels can be transferred to allow developmental flexibility. Therefore, the gross FAR was used as a measure of economic performance. Based on the zoning limits, the configuration of a typical superblock and the design schema, the gross FAR limits of the peripheral and interior region, were calculated as 6.6 and 2.5, respectively, which were used as the peripheral and interior FAR limits. A normalized one-tail FAR tolerance of 0.003 was introduced to mimic the real urban development situation. The objective function assesses the total deviation of two FAR values from the zoning limits with the given tolerance, and the optimization problem was defined as (equations (1) to (3))

minimize : f_{1} (S) = D_{p} (S) + D_{i} (S)

(1)

with

D_{p} (S) = \max (\frac{FA R_{p} (S)}{6.6} - 1, 0, 1-0.003- \frac{FA R_{p} (S)}{6.6})

(2)

D i (S) = \max (\frac{FA R_{i} (S)}{2.5} - 1, 0, 1-0.003- \frac{F A R_{i} (S)}{2.5})

(3)

where:

S - design scenario defined with the design schema and parameters

D_p - function to calculate the deviation of FAR from the zoning limit in the peripheral area

D_i - function to calculate the deviation of FAR from the zoning limit in the interior area

FAR_p - function to calculate the FAR of the peripheral area

FAR_i - function to calculate the FAR of the interior area

Representation update

Based on domain knowledge of urban design, it was argued that the initial design schema is too general producing unrealistic urban forms; for example, infeasible size and shape of buildings were generated and impossible street intersections emerged. To address this problem, a series of constraints were introduced such as a minimum distance between street intersections, minimum areas of parcels and buildings, and minimum widths of development parcels. The complete list of design constraints applied in this study to update the representation is given in the online supplemental Table 3.

Optimization with GAs versus stochastic enumeration

In this case study, the optimization method was the GAs, one of the common evolutionary algorithms inspired by the natural selection (Holland, 1992). In GAs, an individual solution is defined with chromosomes to record all parameters, and a group of individuals is called a generation. Best individuals are selected from the current generation to produce new individuals through randomized crossover and mutation operations to form a new generation. Such an iteration continues to find near-optimal solutions (Herrera et al., 1998). Compared to gradient-based and brute force methods as two common optimization techniques, GAs are robust and powerful for difficult problems with the design space that is large, complex, discontinuous, and poorly understood, all very common in urban form optimization (Herrera et al., 1998; Krish, 2011; Li et al., 2017; Oxman, 2008). Such a method has been widely used in engineering design and architectural system design (Gen and Cheng, 2000; Li et al., 2017). However, their use in urban design is rare because of the complexity of urban form and the lack of urban-scale performance tools.

The GAs used in this study adopted a mixed-coded chromosomes representation (Herrera et al., 1998), which contain 26 general parameters in the real-coded part and 300 parameters for the number of floors in the binary-coded part. The selection of operators and hyperparameters is based on the literature of performance comparisons (Deb and Goyal, 1996), including Deb’s sorting and crossover methods (Deb, 2000; Deb and Goyal, 1996), the tournament selection method (Herrera et al., 1998), and the flipping and bounded polynomial mutation methods (Herrera et al., 1998).

In this study, designs were generated and selected iteratively with GAs to minimize the deviation from the FAR limits. Through the optimization with 60 generations with 2000 individuals per generation, a total of 120,000 design variations were generated and evaluated to find 2351 optimal design schemes which reach FAR limits (Figure 3(c)). Generally, the average deviation from the FAR limits decreases and the number of feasible solutions increases through generations (Figure 3(a) and (b)). Based on expert knowledge and local knowledge encoded into the framework, the resulting optimal solutions are potentially feasible design schemes that could emerge from the basic superblock structure as long-term development patterns.

Figure 3.

The optimization results in generations using the GAs: (a) deviation from FAR limits, (b) number of optimal designs variations and (c) axonometric projections of four generated sample designs. GAs: genetic algorithms; FAR: floor area ratio; gen: generation.

To test the efficiency of the optimization with GAs, the stochastic enumeration method that randomly generates design variations was employed as a comparison. The experiment generated approximately 6,000,000 random design variations and only four were found to be optimal. The comparison suggests that the GAs outperformed the stochastic enumeration method by approximately 30,000 times in terms of searching efficiency, significantly better than random search.

Intermediate decision-making

The found optimal designs were presented to designers to encompass a wider possibility of urban form. A sample set of four feasible designs are shown in Figure 3(c). The main concern raised by the design participants was that generating a large number of designs based on FAR goals is far from enough in real design practice. Although FAR is an important objective, other objectives should also be considered to promote sustainable development in Gangnam. That issue guided the next design iteration.

The second iteration: Adding a new objective

Objective and representation updates

With the new goal of improving energy efficiency, the objective function was redefined to estimate the average building energy use intensity (EUI), and a constraint was established that the FAR of the site reaches its zoning limit. The previous objective function f₁ in the first iteration was then used in the constraint equation. The optimization problem was redefined as (equations (4) and (5))

minimize: f_{2} (S) = \frac{EU(S)}{A(S)}

(4)

subject to:

f_{1} (S) = 0

(5)

where:

S - design scenario defined with the design schema and parameters

EU - function to calculate total building annual energy use

A - function to calculate building total floor area

The objective function f₂ was based on the urban scale building energy modeling system developed and validated in previous studies (Quan et al., 2015). This system is a physically-based simulation tool to estimate building energy use at the urban scale based on the geometry and properties of buildings. It uses the EPC (energy performance coefficient) calculator as the core energy simulation engine (Lee et al., 2013) and considers shading and microclimate effects using the Radiance and UWG (Urban Weather Generator) developed based on TEB (Town Energy Balance) as its sub-engines (Ward, 1994; Bueno et al., 2013). The annual energy use of each building was simulated and the average EUI of the entire superblock was calculated as the energy performance index.

The representation was also updated. Building energy use depends not only on building geometry but also on the building material, heating, ventilation, and air conditioning HVAC system, human behavior, and more (Ratti et al., 2005). For simplicity, the new representation adopted the reference building approach. Reference buildings refer to building models typical for certain building types such as residential and office buildings, which include detailed building information particularly for building energy simulation. One of the well-maintained reference building datasets is U.S. Department of Energy reference building dataset developed and maintained by the U.S. Department of Energy based on large-scale surveys and analysis (US Department of Energy, 2011). While such a database is lacking in Seoul, the material and system in new building constructions in Seoul are similar to those in the reference building dataset for New York city. Therefore, in this demonstrative case study, New York’s reference building is used as the best available data to provide insights into building energy performance in Seoul.

Simulation and optimization integration

GAs and the building energy modeling tool were integrated for design generation. Incorporating the complex modeling system with GAs introduces a big challenge in computation because the GAs requires a large number of simulations, and each simulation is very time-consuming. Some tactics were applied to reduce computation time. The first was to skip the energy simulation of the infeasible designs that did not reach the FAR limits. The second was to index the already simulated designs and store the simulation results for reuse when the design emerged again in later generations. The third was to develop graphics processing unit (GPU) parallel computing to simulate all buildings in the superblock for each design. The fourth was to apply distributed computing to simulate several designs at the same time exploiting the parallelism of GAs. These tactics together reduced the average simulation time for each design by 20–30 times.

With the same settings for the GAs as in the previous iteration, the optimization was run twice to prevent getting stuck early at a local optimum. In both runs, the result shows that the designs progressed toward lower EUI through the generations (Figure 4). Although the optimization paths differ, both runs reach the designs with the EUI of 298.6 kWh/m² after 60 generations, a significant improvement from the EUI of 327.9 kWh/m² and 323.6 kWh/m² at the beginning.

Figure 4.

Optimization results from two runs: (a) the first run and (b) the second run. GAs: genetic algorithms; EUI: energy use intensity; gen: generation.

Final decision-making

The optimization intermediate and final results in the process of the two runs at generations 0, 20, 40, and 60 are presented in the forms of plans and axonometric projections in Figure 5. Tracing the optimization process help stakeholders to better understand how the designs have evolved over the generations guided by AI algorithms.

Figure 5.

Two series of plans and axonometric projections of the designs at generations 0, 20, 40, and 60 tracing the optimization processes in: (a) the first run and (b) the second run. EUI: energy use intensity; gen: generation.

In the decision model, the design solutions from the two runs were evaluated and discussed. The urban forms of the two solutions differ in the details despite the identical EUI and their similar structural pattern. To compare these solutions, other objectives such as walkability, daylight availability, and outdoor comfort that also take important roles in urban design need to be considered. More iterations are expected to reframe the design problem for these aspects to refine solutions with opinions from stakeholders such as the public who will be the users of the space. Such iteration can continue to provide more comprehensive and refined solutions. However, due to time and budget limits in design, after the second iterations, a manual judgment was used to choose from the two solutions. With another design objective as the livability of cities, small block size was considered as an additional criterion, as suggested by Jane Jacobs (1961) that homogenous and large urban blocks create boring and dangerous street environments. It was observed that the design solution from the first run is composed of generally smaller buildings and shows an obvious linear structure which can have better vitality of the street where small buildings can open stores on their ground floors. The design layout from the second run, in comparison, seems too rigid and consists of large block sizes. Therefore, the design from the first run was favored out of the two solutions.

System implementation

The Smart Design framework is implemented as a loosely coupled computing system written in Python. The objectives and representations are encoded and managed in an object-oriented database. The generative model and optimization model are developed as separate modules. For the decision-making process, visualizations are also produced in Python. Parallel GPU computing and distributed computing methods are developed for this framework and facilitate the significant decrease in computation time. More specifically, the methods use the Accelerad (Jones and Reinhart, 2017) and official multiprocessing library of Python. With the distributed computing system developed for this study utilizing three i7 workstations, the first and second iterations took approximately 10 hours and 70 hours, respectively.

Conclusions

Planning and design play an important role in shaping our cities, but their complex process remains a challenge in practice. Coincidently, these two concepts are also important topics in AI research, though with more general meanings. What is shared between the studies on them in two distinctive fields is the same interest in knowing how a design decision can be made for a particular situation with computational support. This shared interest provides an opportunity to integrate the knowledge and tools from the two disciplines to tackle the big challenge: how to better support the design for sustainable urban development in a complex urban environment?

Exploration of two sub-questions is necessary to answer this big question. The first one is how the information, knowledge, and tools from different sources and participants can be organized and integrated into a general formal design process within which design problems and solutions co-evolve. The second is how the detailed design scenarios can be generated in the narrow design generation process given well-defined design problems. In the literature, while the PSS focus on the first sub-question, their answers to the second one are rather simplified and static. On the contrary, the generative design school develops sophisticated methods for narrow design generation, but their consideration of the entire design process is mostly linear and even incomplete. Although both schools have been embracing “Science for Design” with the utilization of scientific modeling and simulation, their considerations of “Design in Science” focus only on one of the two types of design processes.

Interestingly, the sub-questions are also addressed in AI research with two types of AI: systems that think like humans, such as knowledge-based AI for the general design process, and systems that think rationally, such as optimization systems for narrow design generation. Although there have been great achievements for automating narrow design process, the attempts to automate the general design process such as the development of expert systems failed largely because of the nearly infinite amount of design knowledge, the complexity of the cognitive system for design, and the difficulty to automate the participation from stakeholders (Batty, 2018; Goel and Davies, 2011). In the recent new wave of AI enthusiasm, those obstacles still remain, despite the increasing ability of AI. A universal AI that automates the entire design process would require considerable technological advances if could ever be realized.

This research developed the Smart Design framework to better address the two sub-questions with the aid of AI to enhance urban design decision-making. Learning from the idea of complex design processes in PSS, the Smart Design framework streamlines the general design process as a human–AI coupling cognitive system that has four stages: the human problem initialization, human–system interface, system optimization, and human–system interaction. Such recursive, dynamic, and participatory structure shares similar ideas in cognitive systems in AI, but it is developed as a coupled system between humans and computers to deal with the complexity of design, as Design in Science. The narrow design stages of the human–system interface and system optimization adopted the AIAD approach with advanced AI search techniques often used in the advanced generative design methods. The AIAD approach and scientific models and tools can help to search for optimal designs efficiently in a huge design space given a particular design subtask for sustainable development, reflecting Science for Design (Batty, 2013b). The Smart Design framework provides a formal structure to connect the two design processes for better urban design decision support with the aid of AI. Each component in the framework including the goals, constraints, parametric representations, methods, and participants can be defined depending on the design problems and context in real design practice.

The Smart Design framework was applied to Gangnam superblocks in Seoul as a case study to design sustainable urban forms. Two iterations were demonstrated to explore the co-evolution of design problems and solutions, and to identify designs with least EUIs and highest FARs as the basis for further decision-making. The case study demonstrates how the framework can be realized with specific components for a particular design case, and how the formal process in the framework can assist the designers and planners to make intelligent design decisions with respect to the dynamically changing objectives for sustainable city development.

The new framework provides a new and powerful way of integrating design, science, and AI for planning and design. The key aspects and features of the framework—including its iterative and participatory aspects, cross-scale structured representations of urban forms, and integration of simulation tools and GAs—are expected to contribute to the re-emerging discussion on AI and design. However, many challenging issues remain in the current implementation of the framework: how to explore and update design schemata; how to decide whether the generative model should use a symbolic or visual representation; how to speedup complex urban modeling tools for dynamically changing objectives; how to apply multi-objective optimizations; and how to manage uncertainties in the process. These issues will be studied in future research.

Supplemental Material

Supplemental material for Artificial intelligence-aided design: Smart Design for sustainable city development

Supplemental Material for Artificial intelligence-aided design: Smart Design for sustainable city development by Steven Jige Quan, James Park, Athanassios Economou and Sugie Lee in EPB: Urban Analytics and City Science

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Creative-Pioneering Researchers Program through Seoul National University (SNU).

Steven Jige Quan is an Assistant Professor in the Graduate School of Environmental Studies at Seoul National University, Republic of Korea. His research fields include energy resilient urban design, AI-aided design, urban energy modelling, and PSS. Steven worked as a chief planner in the No.3 Department of Comprehensive Planning at Beijing Tsinghua Urban Planning & Design Institute, China. Steven holds a B.Arch and M.Eng in Urban Planning from Tsinghua University, China, and a PhD in City and Regional Planning from Georgia Institute of Technology.

James Park is a PhD Student in the School of Architecture at Georgia Institute of Technology. His research investigates courthouses as a building type and focuses on designing and implementing a generative description of the building type in the form of a shape grammar. James is a research assistant at the Shape Computation Lab, an instructor for the series of courses on computational media and modeling at the School of Architecture and a teaching assistant for the Georgia Tech Architectonics in Greece and Italy Study Abroad Program. James holds a BSArch and an MArch from Georgia Institute of Technology.

Athanassios Economou is a Professor of Architecture and Director of the Shape Computation Lab in the College of Design at Georgia Institute of Technology. Dr. Economou’s teaching and research are in the areas of shape grammars, computational design, formal specification of style, computer-aided design and design theory with over forty published papers in these areas. Dr. Economou holds a Diploma in Architecture from NTUA, Athens, Greece, an M.Arch from USC, and a PhD in Architecture from UCLA.

Sugie Lee is a Professor at the Department of Urban Planning & Engineering in the College of Engineering, Hanyang University in Seoul, Korea. He does research in urban form, travel behavior, mobility, walkability, and urban thermal environment. He holds a PhD at the School of City and Regional Planning, Georgia Institute of Technology.

References

Alonso

Zhang

Grignard

, et al. (2018) CityScope: A data-driven interactive simulation tool for urban design. Use case Volpe. In: International conference on complex systems, Cambridge, MA, USA, 22--27 July, pp. 253–261. Berlin, Germany: Springer.

Batty

(2003) Planning support systems: Techniques that are driving planning. In: Geertman

Stillwell

(eds) Planning Support Systems in Practice. Heidelberg: Springer, pp. v–viii.

Batty

(2013a) Defining geodesign (= GIS plus design?). Environment and Planning B-Planning & Design 40(1): 1–2.

Batty

(2013b) The New Science of Cities. Cambridge: The MIT Press.

Batty

(2018) Artificial intelligence and smart cities. Environment and Planning B: Urban Analytics and City Science 45(1): 3–6.

Batty

Yeh

(1991) The promise of expert systems for urban planning. Computers, Environment and Urban Systems 15(3): 101–108.

Buchanan

(1992) Wicked problems in design thinking. Design Issues 8(2): 5–21.

Bueno

Norford

Hidalgo

, et al. (2013) The urban weather generator. Journal of Building Performance Simulation 6(4): 269–281.

Bulkeley

Coenen

Frantzeskaki

, et al. (2016) Urban living labs: Governing urban sustainability transitions. Current Opinion in Environmental Sustainability 22: 13–17.

10.

Campbell

(1996) Green cities, growing cities, just cities? Urban planning and the contradictions of sustainable development. Journal of the American Planning Association 62(3): 296–312.

11.

Chase

(2005) Generative design tools for novice designers: Issues for selection. Automation in Construction 14(6): 689–698.

12.

Colgan

Spence

Rankin

(1991) Steering automated design. In: Gero

(ed) Artificial Intelligence in Design '91. Oxford: Butterworth-Heinemann, pp. 211–230.

13.

Deb

(2000) An efficient constraint handling method for genetic algorithms. Computer Methods in Applied Mechanics and Engineering 186(2–4): 311–338.

14.

Deb

Goyal

(1996) A combined genetic adaptive search (GeneAS) for engineering design. Computer Science and informatics 26: 30–45.

15.

Geertman

Ferreira

Goodspeed

, et al. (2015) Introduction to ‘planning support systems and smart cities’. In: Geertman

Ferreira

J, Jr

Goodspeed

, et al. (eds) Planning Support Systems and Smart Cities. Cham: Springer, pp. 1–17.

16.

Gen

Cheng

(2000) Genetic Algorithms and Engineering Optimization. New York: John Wiley & Sons.

17.

Gero

(1991) Artificial Intelligence in Design'91. Oxford: Butterworth-Heinemann.

18.

Goel

Davies

(2011) Artificial intelligence. In: Sternberg

Kaufman

(eds) The Cambridge Handbook of Intelligence. Cambridge: Cambridge University Press.

19.

Goodchild

(2010) Towards geodesign: Repurposing cartography and GIS? Cartographic Perspectives 66.

20.

Graham

Marvin

(2002) Splintering Urbanism: Networked Infrastructures, Technological Mobilities and the Urban Condition. Abingdon: Routledge.

21.

Herrera

Lozano

Verdegay

(1998) Tackling real-coded genetic algorithms: Operators and tools for behavioural analysis. Artificial Intelligence Review 12(4): 265–319.

22.

Holland

(1992) Genetic algorithms. Scientific American 267(1): 66–73.

23.

Hopkins

(1999) Structure of a planning support system for urban development. Environment and Planning B-Planning & Design 26(3): 333–343.

24.

Jacobs

(1961) The Death and Life of Great American Cities. New York: Random House LLC.

25.

Janssen

(2006) A generative evolutionary design method. Digital Creativity 17(01): 49–63.

26.

Jones

Reinhart

(2017) Experimental validation of ray tracing as a means of image-based visual discomfort prediction. Building and Environment 113: 131–150.

27.

Kim

K-J

Choe

S-C

(2011) In search of sustainable urban form for Seoul. In: Sorensen A and Okata J (eds) Megacities. Tokyo: Springer, pp. 43–65.

28.

Klosterman

(1997) Planning support systems: A new perspective on computer-aided planning. Journal of Planning Education and Research 17(1): 45–54.

29.

Knight

Stiny

(2001) Classical and non-classical computation. ARQ: Architectural Research Quarterly 5(4): 355–372.

30.

Krish

(2011) A practical generative design method. Computer-Aided Design 43(1): 88–100.

31.

Lee

S-H

Zhao

Augenbroe

(2013) The use of normative energy calculation beyond building performance rating. Journal of Building Performance Simulation 6(4): 282–292.

32.

Shao

Zuo

, et al. (2017) Genetic algorithm for building optimization: State-of-the-art survey. In: 9th international conference on machine learning and computing. Singapore, 24--26 February, pp. 205–210.

33.

Lynch

(1981) Good City Form. Cambridge: MIT Press.

34.

Martins

TAL

Adolphe

Bastos

LEG

(2014) From solar constraints to urban design opportunities: Optimization of built form typologies in a Brazilian tropical city. Energy and Buildings 76: 43–56.

35.

Niitamo

V-P

Kulkki

Eriksson

, et al. (2006) State-of-the-art and good practice in the field of living labs. In: 2006 IEEE international technology management conference (ICE). IEEE, Milan, Italy, 26--28 June, pp. 1–8.

36.

Oxman

(2006) Theory and design in the first digital age. Design Studies 27(3): 229–265.

37.

Oxman

(2008) Performance-based design: Current practices and research issues. International Journal of Architectural Computing 6(1): 1–17.

38.

Oxman

(2009) Performative design: A performance-based model of digital architectural design. Environment and Planning B-Planning & Design 36(6): 1026–1037.

39.

Peponis

Park

Feng

(2016) The city as interface of scales: Gangnam urbanism. In: Kim

S-H

Cinn

Ahn

, et al. (eds) The FAR Game: Constraints Sparking Creativity. Seoul: SPACE Books, pp. 102–111.

40.

Quan

Augenbroe

, et al. (2015) Urban data and building energy modeling: A GIS-based urban building energy modeling system using the urban-EPC engine. In: Geertman

Ferreira

Goodspeed

, et al. (eds) Planning Support Systems and Smart Cities. Cham: Springer, pp. 447–469.

41.

Quan

Minter

Yang

PP-J

(2013) A GIS-based performance metrics for designing a low energy urban agriculture system. In: Geertman

Toppen

Stillwell

(eds) Planning Support Systems for Sustainable Urban Development. Heidelberg: Springer, pp. 225–247.

42.

Ratti

Baker

Steemers

(2005) Energy consumption and urban texture. Energy and Buildings 37(7): 762–776.

43.

Rittel

HWJ

Webber

(1973) Dilemmas in a general theory of planning. Policy Sciences 4(2): 155–169.

44.

Rowe

(1987) Design Thinking. Cambridge: MIT Press.

45.

Russell

Norvig

(2016) Artificial Intelligence: A Modern Approach: Kuala Lumpur: Pearson Education Limited.

46.

Schüler

Cajot

Peter

, et al. (2017) The optimum is not the goal: Capturing the decision space for the planning of new neighborhoods. Frontiers in Built Environment 3: 76.

47.

Seoul Metropolitan Government (2016) Seoul urban planning. Report, Seoul Metropolitan Government, Republic of Korea.

48.

Seoul Metropolitan Government (2018) Seoul Metropolitan Government Ordinance on Urban Planning. Ordinance, Seoul Metropolitan Government, Republic of Korea.

49.

Shea

Aish

Gourtovaia

(2005) Towards integrated performance-driven generative design tools. Automation in Construction 14(2): 253–264.

50.

Simon

(1969) The Sciences of the Artiﬁcial. Cambridge: MIT Press.

51.

Steinitz

(2012) A Framework for Geodesign: Changing Geography by Design. Redlands: Esri Press.

52.

US Department of Energy (2011) Commercial reference buildings. Report, US Department of Energy, USA.

53.

Vonk

Geertman

Schot

(2007) A SWOT analysis of planning support systems. Environment and Planning A 39(7): 1699–1714.

54.

Ward

(1994) The RADIANCE lighting simulation and rendering system. In: 21st international conference on computer graphics and interactive techniques. Orlando, USA, 24--29 July, pp. 459–472.

55.

Yuh

O-K

Ahn

K-H

(2006) A study on the development and growth of YoungDong new town. Journal of the Korean Regional Development Association 18(4): 161–182.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.36 MB