Generative design experiments with artificial intelligence: reinterpretation of shape grammar

Abstract

Purpose

“Can artificial intelligence produce architectural plan schemes?” discussion is the starting point of this study. The aim of this paper is to question whether this will be a new method in architectural design by producing plans with artificial intelligence interfaces working with human–computer interaction and to create a discussion environment.

Design/methodology/approach

The main research topic is the evaluation of architectural design decisions with the text-to-image generation AI algorithms method based on shape grammar rules. First, a sample space consisting of Palladio plans or plan diagrams was created. Plan diagram production experiments were made with different interfaces (Midjourney, Dall-e2, Stable Diffusion, Craiyon, Nightcafe), and alternative plan diagrams were recorded as outputs. The discussion of the outputs has been made over architectural design and space.

Findings

In the conceptual design phase of the architectural discipline and in the production of architectural plan scheme, AI algorithms are trending. This interaction imposes a new responsibility on architects. AI can create paradigm shifts in architectural processes with its tools with high data processing potential. On the other hand, in this study, it is emphasized that architecture is not just an act of producing visuals, but a functional act of producing visuals.

Originality/value

The technology is effective in producing architectural plans and directing them to artificial intelligence algorithms. With this study, multi-alternative architectural plan productions were tried with text-to-image bots with fast results. In this direction, a new method proposal has been developed for the conceptual design phase in architecture.

Citation

Çelik, T. (2024), "Generative design experiments with artificial intelligence: reinterpretation of shape grammar", Open House International, Vol. 49 No. 5, pp. 822-842. https://doi.org/10.1108/OHI-04-2023-0079

1. Introduction

Artificial intelligence computer programs can mimic certain aspects of architecture (The program can learn from the language of architects, to discover architectural plans, determine material constraints, customer comments, etc. and information about architecture.). They are potential aids for design; they can act as smart extensions of tool-like programs. (Pask, G., 1969).

For the studies carried out in the discipline of architecture, the production of architectural plan schemes by using the technology of that period in every period has been a motivation. Today, the effect of technology and artificial intelligence in architectural plans production and artificial intelligence algorithms are in question. There are studies in which interfaces and codes are written on this subject, but with this study, plan/plan scheme production was tried with the technology of producing text-to-image. In this direction, different spatial plan schemes have been tried to be created by using artificial intelligence and generative design methods. According to Alexander (1967), although the form in the design has the potential to be produced symbolically, it is not possible to evaluate the architectural production outputs with symbolic methods (Alexander, 1967). For this reason, the Palladian grammar of the multiple alternative plan schemes obtained in certain interfaces by following the method of the study and the spatial evaluation are not only measuring the working efficiency of the algorithms of the interfaces, but also the evaluation of the architectural plan configurations of the outputs. For this reason, the formal and functional use of the outputs of the experiments in architectural design is discussed. In this direction, a new method of generate the architectural plan proposal will be developed for the discipline of architecture.

In this study, artificial intelligence bots which are text-to-image generation bots (Midjourney, DALL-E2, Craiyon, Stable Diffusion, Nightcafe) are to be used as a tool with the aim of producing architectural plan schemes. These bots work with text-to-image generation technology, transforming textual descriptions into images with a semantic setup comparable to the text. All of the bots selected in this study are current and cutting-edge examples of machine learning that generates images from text. Architectural plan schemes of an architect were chosen as an example so that the algorithm, which has text-to-image production technology, can offer alternatives by reaching correct generalizations. This sample space was determined as Andrea Palladio plan shemes. The reason why Andrea Palladio plans, which are rule-based, are used is that they have a symmetrical, rule-based setup in the plan layout. The rules of this systematic editing are explained by Stiny and Mitchell (1978) with the name Palladian grammar (Stiny and Mitchell, 1978). It is thought that Palladio's architectural approach is rule-based, which will both increase the number of data in alternatives and facilitate the evaluation of production outputs by comparing them with the sample.

The method to be used in the study is discussed in stages. First of all, literature research on generative design, shape grammar and Palladian grammar was made, and a sample space consisting of Palladio plans or plan schemes was created simultaneously. Afterward, the plan scheme production experiments with the interfaces to be used (Midjourney, DALL-E2, Craiyon, Stable Diffusion, Nightcafe) were started by testing the algorithms. Plan schemes production experiments were made with different interfaces; alternative plan schemes images were taken as output, and these images were evaluated. If the evaluation methods focus on visual quality, according to resolution and formal similarity, and the rules of space production (Palladian grammar), the production outputs are examined by evaluating the architectural plan qualities (Figure 1).

In today's world, where artificial intelligence has replaced contemporary art and design practices and designs are made through autonomous productive interfaces, production with artificial intelligence has been opened to the discipline of architecture as a conceptual discussion area. As a new agenda for architecture and a field where the possible is discussed, the effects of the digital environment on architecture constitute the main motivation of this study.

2. Shape grammar and Palladian architecture

Architecture contains grammatical rules that make up architectural languages and styles. Accordingly, the main expression tools of architecture, which is a design act, are points, lines and three-dimensional geometric objects (Tok, 2008). Language uses separate components (words), rules (grammar), while components of architecture are architectural forms or compositional techniques chosen from within an existing typical style. Syntax in architectural design involves assembling architectural components with conventions gained from past experience (Özbek, 2004).

Shape grammars were introduced as a way of describing and even creating design languages, one of the fundamental concepts of design and architecture, which can be defined with a certain degree of algorithmic structure in accordance with its rule-based structure (Aksoy, 2001; Stiny and Gips, 1972).

Knight says that a complete design definition can be made through the rules of form defined in terms of spatial relations. Form rules are applied to the initial shape, which is a shape made from shapes in the dictionary, and other shapes produced from it, to produce designs. A ‘shape grammar’, which is ultimately created by a set of form-rules and an initial-shape, describes a set of shapes called a ‘designs-language’ (Knight, 1981). Rollo (1995), on the other hand, defined shape grammar as an iconic language tool used to codify the findings of studies on a design language (Rollo, 1995).

The analytical grammar study of the first architectural structure is the Palladian grammar developed for Palladio's villa designs (Stiny and Mitchell, 1978).

The aims of the shape grammar developed by Stiny and Mitchell (1978) can be listed as follows;

1. Clarifying the overall structure underlying Palladio's building designs within the group studied,

2. Presenting criteria for determining whether a building not included in the original sample group is an example of the Palledian style,

3. Transforming Palladio's architectural grammar into a modern, productive form is to develop a generative mechanism to obtain existing examples and new plan examples in this style (Stiny and Mitchell, 1978).

By transferring the rules obtained as a result of the analyses made by Stiny and Mitchell to the digital environment, hypothetical villas as well as the original villas were obtained (Sass, 2007).

Two important advantages are obtained with the analysis shape grammar. First of all, whether two-dimensional or three-dimensional, these studies can be expressed parametrically. The initial shape, the set of rules to be applied, the set of transformations to be applied, their number and reference point are mentioned. The second important advantage is that the projects can be deformed, restructured or personalized with small parametric changes. Therefore, analyzes create a suitable environment for making designs that can be shaped according to the user's wishes.

The features of design languages ​​with algorithmic structure, which have repetitive features, can be likened to the rules of the language. With the analysis of these rules, new forms can be produced with the original language and rules (Stiny, 1980). In this study, the production of new architectural plan alternatives based on Palladian architecture rules was tried by using artificial intelligence technology's own learning and production skills.

3. Generative design experiments with artificial intelligence: reinterpretation of shape grammar

In the syntactic sense, design can be defined as a combination of forms and relations between forms. Shape grammar consists of a shape repertoire, shape rules that define spatial relations, and the initial form in the shape repertoire (Stiny and Gips, 1972; Wojtowicz and Fawcett, 1986). Architectural plan scheme production problem has been the subject of many researches in the literature. Eastman (1973) defined architectural plan-scheme layout problems as space planning problems and said that he aimed to solve this problem with automatic production (Çelik, 2023b; Eastmen, 1973). Between 1955 and 2020, 50 studies were conducted on applications that produce automatic plan layouts in the space planning problem (Uzun, 2020). The first study on this subject was Buffa's (1955) “sequence analysis for functional layouts” (Buffa, 1955).

In the 1970s, it was possible to produce and derive plan diagrams using shape grammar. In addition, starting from the same years, genetic algorithms were also tried in the production of the plan scheme. Today, it is observed that today's artificial intelligence technique deep learning method has started to be used in studies related to the production of plan charts. Krejcirik (1969) developed his studies on optimum settlement in architectural plan with computer-aided design (Krejcirik, 1969). Weinzapfel et al. (1971) geometrically systematized spaces in order to produce spatial arrangements in 3D in the computer-aided design process and used the relationships between these geometries as design parameters (Weinzapfel et al., 1971). Levin (1964) and Grason (1971) developed graph theory for spatial fiction in the plan scheme (Levin, 1964; Grason, 1971). Eastman (1973) proposed a system for automating spatial planning in two dimensions in his research. Eastman (1973) defined three variables for the automatic plan scheme design system: Space, relationships between design units, and operators. It has been said that the operators manipulating the design from these variables are not well defined and provide the potential for the creative point in the design (Eastmen, 1973).

This large number of studies on the architectural plan scheme production problem has led to the diversification of the methods applied in the solution of the problem (Çelik, 2023b; Uzun, 2020). It is seen that the use of deep learning method, which is one of today's artificial intelligence techniques, has started in studies for the production of autonomous plan scheme (Figure 2).

In general terms, being “productive” is characterized as having the power to create or showing the source of this power (Fischer and Herr, 2001). “Productive design”, on the other hand, can be defined as the method in which the performer/designer deals with the process and the content of the process rather than the result, and the “productive design system” can be defined as the system that supports the user in this process or handles the design completely (Fischer and Herr, 2001). The term generative art is now being asked again, “What is art?”; it has created a new space for itself over time. While the fields of architecture, music and industrial design contain examples of productive systems, visual arts also define their own rule-based and functional autonomous systems (Galanter, 2016).

This interest in digital production has also prompted a broad shift in theoretical concerns. If the 1980s and 1990s were characterised by an interest in literary theory and continental philosophy – from the Structuralist logic that informed the early Postmodernist quest for semiological concerns in writers from Charles Jencks to Robert Venturi, to the post-Structuralist enquiries into meaning in the work of Jacques Derrida that informed the work of Peter Eisenman and others – the first decade of the 21st century can be characterised by an increasing interest in scientific discourses. It is as though the dominant logic of today has become one of technology behaviour. (Leach, 2009)

Today, with the new tools of technology, new methods are being tried for the production of architectural plans in the context of “productive design”. One of the new tools is artificial intelligence algorithms, which is the current subject of technology. Bots that produce images from text, working with machine learning technology, are one of these technologies. Machine learning is an application of artificial intelligence that gives computers the capacity to learn automatically from experience and evolve without being programmed (He and Deng, 2017). Text-to-image rendering or generation refers to computer approaches that transform textual descriptions into images with a semantic structure similar to text. In tasks such as image classification, the content of an image is usually uncomplicated with a distinct element to classify. The problem becomes much more difficult if computers are asked to comprehend complex scenarios (He and Deng, 2017). Today, there are “bots” that transform complex scenarios into images on various digital platforms. In informatics terminology, algorithms that perform a task for a specialized purpose online are called “bots”. Bots with the ability to recognize and classify images are one of the new achievements in machine learning (Goodfellow et al., 2014; Zhang et al., 2019).

“Can artificial intelligence/machines produce architectural plan schemes?” as a sub-question of the question “Can artificial intelligence/machines design?” controversy is the main concern of this study. In this context, different spatial plan schemes have been experimented in this study by using generative design methods and artificial intelligence. “Midjourney, DALL-E2, Craiyon, Stable Diffusion, Nightcafe” tools from AI bots were used as tools to autonomous architectural plan generation. With purpose to suggest multiple alternatives by reaching the correct generalizations of the algorithms, the plan schemes of a single architect as Andrea Palladio plan schemes, which was explained by Stiny and Mitchell (1978) with the name Palladian grammar, were determined as a sample space. The reason for using Andrea Palladio plans is that there is a symmetrical and rule-based setup in the plan layout. It is thought that this systematic construct and rule-based situation will both increase the number of data in alternatives and enable the production outputs to be evaluated by comparing with the sample.

First of all, a topic-based pool (Figure 3) was created for the creation of keywords for artificial intelligence interfaces with machine learning technology that produces images from text. Andrea Palladio plans and plan diagrams (Jupp, 2005) (Figure 4) reached for AI bots that allow visual addition were also used as reference images in the study.

The first experiments (Table 1) were carried out on the Midjourney AI visualization interface. Discord is a platform that provides instant chat, file sharing, voice channel connections by creating communities' own servers. Bots were developed to turn Discord into a more interactive environment, as programs that display autonomous behaviors that appear to be normal users (Verma et al., 2021). The bot software Midjourney works with descriptive word codes following the “/imagine” command. As a result, it allows the selected alternatives to be taken in higher resolution or to produce other alternatives of the selected alternative.

Within the scope of the study’s keyword diagram in the Midjourney text-to-image generation bot, 64 visual outputs were obtained. In almost all of them, symmetrical plan schemes have been reached in accordance with Palladian grammar. It has been observed that the resulting products can be used like “architectural plan scheme” in the conceptual stage of design, but in some results, there are results that can be considered as architectural collages or presentations with high visual quality. In this context, the artificial intelligence bot incorporated images of Palladian architecture. There are also fragments and writings from the views and sections inside/beside/around the plans. For those selected from the four alternatives (Figure 5) provided by the interface, higher resolution image acquisition option was applied.

In the second stage, architectural plan creation experiments were carried out using the DALL-E2 that is also text-based like midjourney and also allows uploading reference images. In the DALL-E2 bot, firstly, using the keywords from the word pool in the study, tests were made for the production of architectural plans (Table 2).

As seen in Table 2, the images obtained are far from the architectural plan or plan scheme, instead, they could not go beyond combining some parts of the architectural plans that are already in the memory of the artificial intelligence. For this reason, attempts to produce images with keywords were not continued. As another method, production was made by giving visual reference. First, Andrea Palladio floor plans, Figure 4, were loaded into the interface and the results in Table 3 were obtained.

The visual quality and resolution of the results obtained are not high, but approximately 128 results were obtained in line with the architectural plans given. It can be said that these can be used as conceptual schemes that can be used in preliminary design in the first phase of the architectural design process. In most of the spatial fictions, which seem to have a certain rule base, there is also symmetry in accordance with Palladian grammar. It is thought that this method, in which autonomous and multiple alternatives are obtained very quickly, can also achieve functional results under the control of the architect, based on or using the alternatives obtained. After this successful trial, another experiment was made by giving visual reference again. This time, Andrea Palladio plan schemes, which are Figure 4, were loaded into DALL-E2 and alternatives were requested from the interface (Figure 6). However, the results were not successful because their resolution was very low.

The plan image of the reference selected structure in the Dall-E2 AI interface was taught to the algorithm visually, and the alternatives were produced and final tests were made in this interface (Table 4). The first image was chosen as Andrea Palladio's Villa Cornaro, and the second example is Palladio's Villa la Rotonda. This experiment gave more successful results than the others. Architectural plans, which are the resulting visual outputs, have the spatial characteristics of the architectural plans given as reference. As a method proposal to architectural design, this stage is evaluated as the most successful results.

Other attempts have been made in Craiyon, the current version of the DALL-E Mini. Craiyon – DALL-E Mini software pools images referenced by word codes from unfiltered internet databases (Çelik, 2023a). At this stage, the key words determined in the study were written in Craiyon and the outputs of the production (Table 5) were obtained.

72 visual outputs were obtained by typing words from the keyword matrix within the scope of the study in the Craiyon bot. In these trials, different alternatives that are thought to be used as architectural plan schemes have been reached. In addition, it is possible to describe the results obtained as architectural collage or architectural presentation technique. Although the architectural plan and plan scheme are given among all keywords, the alternatives with Palladian facades on or next to the plans, which are in the program's own memory, are also taken as the final product.

Other architectural plan generation experiments were made in the Stable Diffusion interface (Table 6). Stable Diffusion, like other artificial intelligence interfaces, is a text-based bot. 16 different alternative results were printed out; however, although there are architectural plans in the given keywords, 3 plan fragments, 1 site plan, 8 facade/facade drawings and 4 perspective, results were obtained. For this reason, although the resolutions of the obtained results are high, the Stable Diffusion interface was found to be unsuccessful in terms of architectural plan generation.

Finally, architectural plan production was tried by giving keywords in the Nightcafe interface, but no successful results were obtained (Table 7).

The visual outputs obtained in the architectural plan production experiments worked on different interfaces have changed in the context of the working principles of the artificial intelligence bot (Figure 7), which is the tool used, both successful/unsuccessful in plan production and stylistically differentiated. To summarize, Midjourney was successful in producing the plan diagram. In particular, the production of symmetrical architectural plans referring to Palladion diagrams, which is the sample space given, is impressive. It is suggested to discuss the functionality of these plan schemes in the spatial context with these plan schemes obtained in future studies. Dall-E2, on the other hand, produced images from text that were far from architectural plans. However, with Dall-E2's ability to produce images from images, qualified architectural plan results could be achieved. These can be used as inspirational plan schemes at the concept stage. But especially when it was given an example plan scheme and asked for an alternative from the machine, results that can be used as purely architectural plans are provided. The functional qualities of these plans are suggested to be the subject of the next study. Although the Craiyon bot produces very fast and very productive productions, it does not give results to the concern of architectural plan production both in terms of visual quality and architectural quality. It can be said that they are results that can be inspired at the concept stage. However, successful results were not obtained in the production of plans from Stable Diffusion and Nightcafe bots. In this context, it is seen that the artificial intelligence bot used in the generation stages of architectural plan/plan schemes has now become a design parameter.

The main motivation of this study was that one of the alternative productions to the production of architectural plans, which is one of the practices in the discipline of architecture and which is still a problem, could be the technology of producing images from text. When the results are examined, it is thought that plans can be produced with this technology, which is still in its infancy, and that it will inspire architects with its fast and multi-alternative results, at least in the concept stages.

4. Discussion

As Schumacher (2011) explains, “architecture is in a cycle of “innovative adaptation”: the experimental and productive scenario was digitally retooled and adapted, with the role to organize and articulate their complexity in order to create a repertoire guided by the same patterns” (Schumacher, 2011). In this context, the transformation of traditional design processes with the active role of the architect, the autonomy of the design processes and the transformation of the “designer” identity into a decision maker are becoming the focus of a current discussion.

One of the practices within the discipline of architecture is the production of architectural plans, which is still a current problem. Shape grammars are one of the methods used for this problem. Today, the new tools of technology offer new methods for the production of architectural plans. In this study, new plan schemes were produced by creating a shape grammar sample with bots using the text-to-image-generation method with human–computer interaction. Considering the production of architectural plans in architectural design in general, the use of text-to-image generation bots can provide both positive and negative repercussions:

Iterative design process: Different alternatives based on iteration were produced with the texts from the selected sample in the study. In this context, with text-to-image rendering bots, architects will be able to quickly create multiple design options and variations based on textual inputs. This will ensure an iterative and exploratory design process.

Efficiency and time savings: Decision-making is a process in which alternatives are evaluated in order to choose an option in order to achieve desired goals. This process begins with the recognition of the problem that needs to be decided first. In architectural design, the problem is based on more than one parameter (Menassa and Baer, 2014). When it is multi-criteria, it becomes more difficult to decide on uncertainty (İlerisoy and Gökgöz, 2022). Automatic rendering can potentially speed up the architectural plan generation process. Creating initial visual representations via bots can provide a starting point for further refinement and reduces the time required to manually create initial design sketches or renders.

Interpretation limitations: Text-to-image generation bots may have difficulties correctly interpreting and translating complex architectural concepts and design intent. They may not fully grasp the context, nuances or spatial relationships that architects envision. In the study, when an architectural plan was requested, some bots produced architectural plans, while others produced architectural forms. In the same way, results were obtained in the form of plan presentation rather than architectural plan quality. As a result, the images created may not always be compatible with the original intention of the architect, leading to potential misunderstandings.

Lack of architect's touch: Architectural design is more than just visual representation. Functionality and context need to be understood in the design processes. As seen in the study, although the architectural plan production is obtained schematically, its architectural functionality should also be tested.

As a result, while text-to-image generation bots can serve as inspiration during the preliminary design/concept phase, enhance visualization and support the architectural design process, they have limitations in fully capturing the design intent and holistic aspects of the architecture. It is important to approach their use as a substitute for human creativity, expertise and critical thinking. In addition to these, when the results of each bot, which are the tools used in the study, are evaluated, it is seen that each bot gives successful or unsuccessful results in line with its own possibilities and abilities. For this reason, it would be correct to consider these artificial intelligence bots not only as a design tool but as a design parameter.

Another issue that needs to be discussed about the findings of the study is whether artificial intelligence can be used in architectural education and whether it should be introduced in design studios. In this regard, Alalouch (2018) said that “It is a necessary pedagogical goal to ensure that educating students in a new way of thinking and discovering process-oriented and they have a strong visual and output-oriented architectural culture.” (Alalouch, 2018) For this reason, it is thought that the relationship between artificial intelligence and architecture, which can be shown as an inspiration in design studios in architectural education, should be the subject of future studies.

5. Conclusion

In the conceptual design phase of the architectural discipline and in the production of architectural plan scheme, artificial intelligence algorithms are trending. As researches involving artificial intelligence and architecture interaction started to increase in the literature; “Is the interaction of artificial intelligence and architecture disciplines necessary?”, “Can artificial intelligence participate in the act of designing algorithms?”, “If artificial intelligence participates in the act of designing, how should the production outputs be evaluated?”, “Is it possible for the profession to become autonomous with artificial intelligence?”, “Artificial intelligence, what should be the future role of the architect with?” questions become important. This interaction imposes a new responsibility on architects. Artificial intelligence can create paradigm shifts in architectural processes with its tools with high data processing potential. We can see that the first attempts pointing to this situation in the discipline of architecture focused on the problem of production of plan schemes. When the Palladio plan schemes, which are the output of this study, are examined, it is emphasized that architecture is not just an act of producing visuals, but a functional act of producing visuals.

Although artificial intelligence algorithms in today's technology achieve productions that can capture visual similarity, the experiments lack full control by including a randomness that should be the responsibility of the architect. The traditional design processes, in which the designer and architect takes an active role, are approaching to become autonomous with the inclusion of collective digital environments in the design processes, and it creates a new focus of discussion as it starts to transform the designer identity into the decision mechanism of among the many alternatives. Digital interfaces work on autonomous design mechanisms that can be described as generative design systems. When the architectural plan or architectural plan scheme visuals, which are the final products obtained within the scope of this study, were examined, some results were unsuccessful and only building visuals were obtained. In addition, although productions evaluated only in terms of formal similarity can be realized in the results of the architectural plan or plan scheme, that are successful results, it would not be correct to say that it is a complete “design” action. Although it is not a complete design action, it is thought that artificial intelligence technology should be evaluated as a new design method or design parameter. For these results, it is seen that multiple alternatives of the architectural plans are obtained autonomously and fast, regardless of the “design” quality for the desired results. It is thought that these results can be considered as a source of inspiration in the “conceptual design stage”, which is the first stage of the design. However, the fact that it will save time in decision-making processes in architectural design and that it should be considered as a parameter in design also shows that artificial intelligence bots have a significant effect on plan production. In this study, shape grammer sampling, which is an architectural plan production method, was used and reinterpreted through technology.

The rapid development in artificial intelligence and technology suggests that in the future, for the discipline of architecture, artificial intelligence may turn into an algorithm that can learn the act of “designing” with the developing new hardware and algorithms. The discipline of architecture will develop and transform together with the developments coming from technology.

Figures

Figure 1

Flowchart of the study

Figure 2

Methods used in architectural plan scheme production studies in the literature

Figure 3

Word pool network diagram within the scope of the study

Figure 4

(a) Andrea Palladio floor plans (b) Andrea Palladio plan schemes

Figure 5

Architectural plan experiments with Midjourney

Figure 6

Architectural plan experiments using reference image on DALL-E2

Figure 7

The working principles of the artificial intelligence bots

Table 1

Architectural plan experiments with Midjourney

Table 2

Architectural plan experiments with DALL-E2

Table 3

Architectural plan experiments using reference image on DALL-E2

Table 4

Architectural plan experiments with a single building reference image in DALL-E2

Table 5

Architectural plan experiments with Craiyon

Table 6

Architectural plan experiments with Stable Diffusion

Table 7

Architectural plan experiments with Nightcafe