1 Introduction
Against the backdrop of ongoing developments in computational technology and artificial intelligence, the field of architectural design is undergoing a profound shift from computer-aided to computation-embedded practices. Traditional workflows centered on drafting and modeling are increasingly being replaced by algorithmic optimization, data-driven analysis, and AI-based generation (
Gaudilliere, 2019;
Caetano and Leitão, 2020). This transformation signals more than just an upgrade in tools and techniques―it reflects a fundamental reconfiguration of design processes, cognitive models, and creative mechanisms.
This shift also brings new challenges to the skillsets, professional roles, and identities of architects (
Oxman, 2008;
Yu et al., 2015). While architects have conventionally led the conceptual and creative aspects of design, and pre-digital modes of ideation―such as hand sketching or conceptual modeling―continue to play a complementary role, the rise of intelligent technologies is redefining the boundaries of architectural practice (
del Campo, 2024). The architect’s authority, expertise, and judgment within the design process are being renegotiated. As
Marković et al. (2024) suggest, architects who are disconnected from digital technologies may soon be as limited as those who once lacked drawing skills. In this rapidly changing computational landscape, architects must actively reposition themselves through self-adaptation and knowledge reconstruction in order to maintain their agency within increasingly digital design systems (
Deutsch, 2017;
Horvath, 2022;
Pouliou et al., 2024).
Existing studies have extensively examined the functional capabilities of emerging technologies and their impacts on design efficiency, creativity, and collaboration (
Chew et al., 2024;
Li et al., 2025). A foundational aspect of this research lies in how computational methods streamline the design process and accelerate decision-making. These approaches typically aim to ensure building performance and compliance with key criteria, enabling rapid iteration and narrowing down solution spaces (
Zhou et al., 2024;
Yan and Liu, 2024). Other studies have explored how computational tools complement the working logic of architects―for instance, by extracting, encoding, and systematizing design knowledge, including precedents and rules, in order to partially automate early-stage decisions (
Cifuentes, 2025;
Zhang and Li, 2025). These approaches have also sparked debates around the risk of over-formalizing design processes, particularly in balancing computational precision with the subjectivity of design act. Some researchers argue that computational tools should serve as provocations for creative exploration, granting architects interpretive agency and open-ended control (
Pouliou et al., 2023). Others propose that structured data and domain knowledge can support new forms of creative expression through AI-driven workflows (
Blaas et al., 2023). Recent discussions have also emphasized the potential for such technologies to rationalize design and building processes, particularly in the service of sustainability and informed decision-making (
Vite et al., 2021;
Stojanovski et al., 2022). Additionally, several studies have addressed emerging modes of human–computer collaboration, focusing on new forms of interaction (
Dade-Robertson, 2013) and data communication (
Shi et al., 2023).
Despite these valuable contributions, comparatively little attention has been paid to the deeper, evolving synergy between architects and computational systems. That is, how architects might engage with the underlying logic of such systems to reconfigure not only outcomes, but also the very processes through which design unfolds. This form of synergy demands a renewed understanding of the relationship between design thinking and computational thinking. As
Wing (2006) defined it, computational thinking is concerned with determining what is computable, using fundamental computational concepts to solve problems, design systems, and understand human behavior. Design thinking is typically grounded in spatial reasoning, precedent-based intuition, and visual judgment, whereas computational thinking introduces new modes of abstraction, decomposition, rule definition, and system construction. These two approaches are not inherently contradictory; the introduction of computational thinking does not imply subordinating design to fixed technical logics. On the contrary, it offers architects alternative lenses through which to perceive and structure the design process (
Calisir Adem and Cagdas, 2020). Their integration holds potential for establishing a more powerful and reflective framework for human–computer synergy in architecture.
This study draws upon the authors’ own experiences in developing a series of web-based generative design tools as a lens to reflect on shifts in architectural roles and the corresponding restructuring of thought processes. These tools are not merely case studies, but active components of the research process, through which critical insights were gained into the dynamics of human–computer collaboration. In this view, architects are not merely tool users, but also developers of generative algorithms, and organizers of workflows. Drawing on four web-based generative design applications―SIMForms, ANYSite, NEXUSpace, and FLEX-Urban―this paper explores how architects establish synergies with computational systems through the intertwined processes of tool-making and technical reasoning. On this basis, the paper introduces a theoretical framework composed of three interrelated mechanisms of human-computer collaboration: (1) reconstructing the architect’s interaction with intuitive media; (2) adopting structured and deductive algorithmic thinking; and (3) organizing the design process as a programmable and systematized architecture.
By articulating this framework, the study aims to offer a conceptual foundation for reimagining human–machine collaboration in architecture. It emphasizes that architects should assume a compound role as tool users, algorithm developers, and workflow organizers―actively participating in shaping the mechanisms of collaboration. In a context where digital and intelligent technologies are becoming deeply embedded in architectural practice, reexamining the nature of human–computer synergy is crucial not only for understanding how tools transform design, but also for envisioning the future competencies of the profession.
2 Seeing as designing: reconstructing intuitive communication
The rapid development and widespread adoption of computational technologies have significantly altered the logic of design. The increasing complexity of algorithmic models, higher levels of automation, and the integration of graphic hardware with AI systems are progressively reducing many design tasks to matters of parameter configuration and logical inference (
Calixto and Croffi, 2024). Particularly in the early conceptual and schematic design stages, mechanisms such as procedural generation, rule-based reasoning, parametric control, and automated optimization are now commonly employed in form-making and performance evaluation. At first glance, design appears to have become “simpler”―powerful algorithms and computational capabilities allow designers to quickly generate feasible solutions in terms of form, structure, and even spatial organization.
However, this perceived simplicity is often superficial, reflecting an acceleration in outcome generation rather than a reduction in cognitive effort. In fact, the complexity of technology has not disappeared―it has merely been reorganized and encapsulated within tools. What matters is whether this encapsulation opens new cognitive pathways or instead disrupts established mechanisms of design judgment. In earlier phases of computational design research and tool development, architects were often required to directly engage with complex mathematical models or abstract data interfaces, sometimes forced to operate within computational logics alien to their design instincts. As a result, key processes of feedback, evaluation, and iterative refinement were compressed―or even broken. While generative results were produced rapidly, the feedback mechanisms became delayed, indirect, or altogether opaque. Designers no longer interacted with a medium that responded intuitively to their intent, but with an input–output-oriented technical system (
Bernal et al., 2015), in which design judgment was often subordinated to the ability to operate and comprehend the underlying tool logic.
SIMForms was developed as a response to the challenges discussed above. Rather than further amplifying technical complexity, it aims to reconstruct a more intuitive, realtime, and visually responsive feedback mechanism by leveraging existing parametric modeling and AI capabilities. The system consists of three main modules (Fig. 1). The first is the form generation module, which incorporates over ten predefined logics for common architectural massing strategies such as shifting, chamfering, and stacking. Users can generate a diverse range of controlled formal variations by adjusting just three fundamental parameters: floor height, setback distance, and expected floor area ratio (FAR). These variations are produced through rule-based procedural algorithms that encapsulate how parameters affect site boundaries and form composition. This encapsulation allows users to focus on understanding the quantitative and geometric relationships between input parameters and resulting forms, rather than the computational details behind them. The second is the performance metrics module, which runs in real time during model generation. By tracking both volumetric and non-volumetric components, it calculates eight key indicators: site area, building height, building density, total floor area, FAR, surface area, shape coefficient, and south-facing façade area. These metrics are displayed immediately in the user interface, enabling designers to rapidly iterate on their schemes based on performance feedback. The third is the AI rendering module, where the 3D model scenes can be automatically captured as input images for an integrated AI renderer based on a pretrained Stable Diffusion model. Eight preset rendering styles are available, allowing users to select a rendering scope and generate stylized visualizations with a single click. Advanced options for fine-tuning model parameters and comparing multiple outputs are also supported.
The primary intention behind the design of SIMForms is not to substitute architectural judgment with computational operations, but rather to reconstruct a form of intuitiveness that aligns more closely with architectural thinking. Instead of passively adapting to the data structures or mathematical paradigms preferred by machines―or relying on conventional modeling or drafting approaches―architects can work from the levels they are most concerned with: formal strategies, index control, and visual representation. SIMForms integrates parametric modeling, performance evaluation, and AI-based image generation into a unified and simplified “form–metric–visual” workflow. This simplification does not imply a dismissal of technical complexity; rather, it seeks to strategically encapsulate that complexity so that critical decision points are exposed to users in a more intuitive and accessible way. Architects are not required to input opaque technical parameters or operate according to the logic of generative algorithms. Instead, they can engage with familiar design controls―such as setback distances, floor heights, and FAR targets―as well as architectural vocabularies involving style, form, and visual qualities. Each design action―be it parameter tuning, form selection, or image generation―is grounded in the architect’s observational and evaluative priorities, making the feedback process more immediate, tangible, and interpretable.
In this sense, SIMForms does not merely pursue a “simplification” of the design process, but redefines what simplification means. It does not hand decision-making over to closed black-box algorithms. Rather, it uses computational power to support architects’ intuition-driven judgments and expressions (
Nisztuk and Myszkowski, 2018), enabling technology to serve architectural thinking, rather than requiring architectural thinking to conform to technological logic.
3 Encoded thinking: structuring architectural knowledge through algorithms
Building on the intuitive design logic exemplified by SIM-Forms, its feedback mechanisms are reoriented toward the architect’s cognitive priorities, enabling a “seeing is designing” approach. However, this pursuit of immediacy and responsiveness also raises a critical question: can such convenience lead to a new form of technological complacency? As architects benefit from the ease brought by computation, are we also increasingly avoiding a more fundamental challenge―how should design thinking be translated into computational language? When the inner workings of computational systems remain concealed behind the interface, can architects still fully understand and direct the generative mechanisms of design?
This concern points to a broader competency increasingly expected of contemporary architects: not only how to use computational tools, but also how to build them. As computational design becomes more pervasive, the boundaries of architectural expertise are expanding. Designers are now required not only to possess disciplinary knowledge in architecture, but also to develop methodological capabilities for embedding their design strategies into computational systems (
Çalışkan‚ 2017;
Yu et al., 2021). In this light, computational systems should no longer be treated as passive black boxes, but rather as externalized representations of architectural cognition. This externalization involves translating design decisions and reasoning into algorithms―it is the articulation of implicit rules and the encoding of design thinking itself (
Menges and Ahlquist, 2011;
Watanabe, 2018). From this perspective, the critical issue is not merely whether architects can code, but whether they can encode their design intentions computationally―constructing design methodologies and operational systems that are executable by machines. The development of ANYSite and NEXUSpace is situated within this context. These tools do not aim to highlight technical complexity; instead, they focus on how design knowledge―ranging from spatial strategies and compositional logic to generative rules―can be embedded into a structured, logic-driven generative framework.
ANYSite is a generative tool designed to produce hierarchical plot subdivision schemes for large-scale urban sites. Its core objective is to generate plot layouts that respond to existing urban textures while maintaining a layered spatial hierarchy (Fig. 2). While real-world plot division is influenced by complex socio-political and environmental factors, ANYSite―as a prototypical tool―focuses on two specific aspects of designers’ reasoning: the relationship between plot geometry and urban fabric, and the control over the number and size of individual plots. To represent urban texture, the system employs a tensor field model derived from the geometric analysis of site boundaries and surrounding building layouts. This model generates a directional vector field reflecting the local urban morphology. Designers can specify subdivision strategies hierarchically, including parameters such as the number of sub-plots and target area ratios. The system then automatically generates subdivision proposals and employs a multi-agent optimization process to adjust plot shapes, ensuring alignment with the site’s spatial logic while achieving more coherent proportions and directional consistency. All subdivision results are encoded in a tree-based hierarchical data structure, enabling users to flexibly view, modify, or revert to any stage of the process.
Building on the capabilities of ANYSite, NEXUSpace addresses a dual challenge in generative design: how to define and manipulate building prototypes under the influence of urban textures, and how to allocate functional spaces efficiently within constrained building volumes (Fig. 3). NEXUSpace also utilizes a tensor field model to capture the spatial characteristics of the site environment, which serves as the basis for shaping building massing. The form-generation process is abstracted through a topological graph model that defines the building’s spatial skeleton. Using an interactive interface, users can sketch out primary circulation paths, which serve as the backbone of the building layout. Subsequently, the system performs automatic plan optimization through a deformation mechanism guided by the tensor field and a spring-based model, ensuring that the resulting geometry is both morphologically coherent and structurally balanced within the site’s context. Massing volumes are then generated alongside each circulation path, with floor height and number of stories defined by the user to produce preliminary 3D building forms. Moreover, NEXUSpace integrates a functional layout module that transforms the space allocation task into a mathematical optimization problem. Designers input a room specification table detailing room types, quantities, area ranges, and other constraints or priorities. Based on this, the system constructs a mixed-integer linear programming (MILP) model, which is solved server-side to yield the optimal allocation. The objective of the optimization is to maximize functional priority matching while minimizing unassigned room units, all within the limits of available building volume. The final layout is visualized and returned to the user, with rooms color-coded by function to support subsequent design evaluation and decision-making.
Although ANYSite and NEXUSpace address different levels of design problems―plot subdivision at the urban scale and massing/layout generation at the architectural scale―they both converge on a fundamental question: How can design thinking be encoded? These two prototype systems do not attempt to replicate the entire workflow of architectural practice. Instead, they focus on encoding those persistent but often implicit cognitive mechanisms in design―such as the understanding of form, the logic of scale control, and the prioritization of spatial organization―into generative systems that can be computationally executed.
ANYSite does not aim to simulate “how urban subdivision should happen” in reality; rather, it asks “how designers perceive and manipulate the logic of subdivision.” Through the use of tensor fields, environmental textures are translated into computable vector information, providing a quantifiable foundation for designers’ intuitions about directionality and morphological continuity. The hierarchical plot generation logic and parameter control mechanisms further transform the implicit awareness of scale into an explicit operational structure. Similarly, NEXUSpace seeks to encode a fundamental tension in architectural layout design: the negotiation between site constraints and spatial intent. On one hand, it uses tensor-guided deformation to adapt building morphology to its surrounding environment at a macro scale. On the other hand, it reconceptualizes the problem of functional layout as a spatial resource optimization task, enabling designers to express and resolve their organizational strategies in computational terms.
The significance of ANYSite and NEXUSpace, therefore, lies in their demonstration of a possible pathway: that design need not rely solely on accumulated intuition or experiential knowledge, but can be advanced through modeling, parameterization, and logical structuring. The act of encoding is not an end in itself, but a mechanism for externalizing design thinking―an essential condition for architects to regain authorship over technological logic. The core issue is not what the tool can do, but how the designer chooses to think.
4 Designing the process: reconfiguring collaborative mechanisms
In the current era where digital technologies are deeply embedded in architectural practice, the term design often refers simultaneously to two intertwined dimensions. On one hand, it denotes the generative process of the design object―how form is shaped, space is organized, and function evolves. On the other hand, it implicates the structuring of the design activity itself―how decisions are made, how designers interact with technology, and how the boundaries of the design process are defined. As
Terzidis (2006) argued, “
If architecture is to embark into the alien world of algorithmic form,
its design methods should also incorporate computational processes.”
It is worth acknowledging that in many contemporary architectural practices, the use of digital tools and the targeted development of computational modules have become increasingly common. As discussed in the previous two sections, their value lies in improving efficiency, assisting creativity, and facilitating the evaluation and resolution of complex problems―thereby enhancing and transforming the design process they serve. Building upon this, we may propose a further line of inquiry: can the design process itself become a reconfigurable system? This implies that architects, by strategically customizing algorithmic logics and interaction mechanisms, may directly reconstruct how the design unfolds, rather than simply enhancing existing processes. Such a shift opens up the potential to challenge habitual modes of design thinking. This form of engagement clearly goes beyond simply “using tools” or defining algorithms for producing architectural outputs. Instead, it moves toward designing the entire computational design system, enabling a higher degree of self-expression (
Aish and Bredella, 2017;
Hirschberg, 2020). Design, in this sense, does not merely occur within tools―it is embedded within the structure of the tools themselves, opening up a new trajectory for exploring how the process of design can be designed.
From this perspective, the development of FLEXUrban aims to respond to this line of thinking through a specific yet illustrative case. It not only builds upon core methods from previous tools in spatial analysis, form generation, and feedback design, but also explores how technical mechanisms can participate in the configuration and redefinition of the design workflow through modular coordination.
The FLEXUrban system comprises two primary modules: an interactive plot subdivision module and a building typology generation module (Fig. 4). In the plot subdivision module, FLEXUrban enhances the previous tensor field approach by introducing a more interactive operation mechanism. All potential subdivision directions are dynamically generated based on tensor field streamlines and previewed in the interface as candidate subdivision lines. Designers can directly select and adjust the positions of these lines and assign street hierarchy levels. This mechanism makes the generation of subdivision lines more aligned with the directional logic of urban fabric, while offering greater operational flexibility and formal control. In contrast to the hierarchical and automated logic in ANYSite, this approach emphasizes the explicit representation of subdivision possibilities and the refinement of decision participation. In the building typology generation module, the system provides 16 predefined building prototypes classified by function and morphology. Based on individual plots, users can select a prototype and customize it through a series of parameters. This process continues the SIMForms logic of intuitive feedback―when users set parameters such as building density, height, or setback, the system immediately generates a 3D model and synchronously presents key design indicators. Newly added parameters, such as maximum overall height and random seed, introduce greater variability and complexity into form control. As a result, the generative process becomes both goal-oriented and capable of producing diverse outcomes.
To further address the repetitive demands of large-scale site design, FLEXUrban introduces an image-driven batch generation mechanism (Fig. 5). The system allows users to upload grayscale images to define parameter fields: in the plot subdivision module, brightness values correspond to the target area proportion of plots; in the building generation module, three separate grayscale images are used to respectively control the expected FAR, building height limits, and buildable areas (i.e., where buildings are placed or left as open space). The system maps pixel brightness values to corresponding parameter values, thereby generating distributed subdivision and building schemes across the entire site. Importantly, batch generation does not preclude user intervention―designers can still adjust subdivision lines and modify building parameters in the generated results, enabling a seamless integration between global control and local editing.
The development of FLEXUrban is not merely a technical integration or functional upgrade; it represents an attempt to align architectural thinking with algorithmic logic, exploring new possibilities for design methodology. For instance, in the plot subdivision module, by externalizing tensor field-based algorithms and underlying data structures into the user interface, the system enables designers to preview and flexibly manipulate subdivision paths generated from different directional logics. This mechanism transforms internal algorithmic inference into a visual and operable interface, where subdivision is no longer a one-off calculation or a generation-verification step. Instead, it becomes a direct design act―completed through algorithmic suggestion, human intervention, and real-time feedback―thus constituting an attempt to reconstruct the design process itself. Particularly noteworthy is the use of image parameter mapping in the batch generation mechanism. By treating grayscale images as input media for parameter fields, designers or planners can use simple image-editing tools to control the subdivision scale and building generation strategies over a large urban area. This mode of interaction can be understood as a visual representation of urban control intentions, especially for urban designers or government stakeholders, offering a new paradigm of human–machine collaboration and decision-making.
What FLEXUrban ultimately gestures toward is a redefinition of the collaborative mechanism between architect and computation. It does not embody a model where technology overrides design, nor one in which designers exercise absolute control. Instead, it proposes a structuring of the design process itself―where the designer acts not only as an operator, but as a rule-maker and logic constructor. Within such a framework, computational tools and technologies begin to define new paradigms for how design unfolds. The computational design process, in turn, becomes an object that can be constructed, configured, and critically discussed (
Gün, 2023).
5 Reimagining synergy between architects and computational systems
Reflecting on the development of the aforementioned web-based design support applications and the theoretical reflections underpinning them, we can distill three key dimensions that characterize the reconfiguration of collaborative mechanisms between architects and computers. These dimensions concern the reconstruction of the architect’s means of operation, modes of thinking, and awareness of design processes:
(1) Designing through intuitive and comprehensible interactions. Architects should be liberated from the mechanical complexities and abstract data structures underlying computational technologies. Instead, they should engage with more immediate, intuitive media of interaction and feedback. SIMForms exemplifies this approach by integrating forms, metrics, and visuals into a unified framework. This establishes a direct and simplified feedback loop, allowing designers to interact with complex systems at a low threshold while maintaining conceptual control in early-stage design.
(2) Thinking through structured and algorithmic logic. ANYSite and NEXUSpace advocate for encoding and structuring design thinking into generative logics. This approach emphasizes the definition of underlying design rules to stimulate creative processes, rather than employing computational tools merely to simulate or approximate preconceived outcomes. Here, architectural thinking shifts toward a more abstract, systematic, and deductive mode, where tacit design judgments are made explicit and translatable into computational strategies.
(3) Orchestrating design through process configurability. Architects not only use tools to support design processes, but also have the potential to shape the structure of those processes themselves. FLEXUrban reflects this shift: rather than simply combining functions, it integrates algorithmic mechanisms and interaction logics to enable a rethinking of how design is initiated, guided, and developed. It restructures the organization of the design process―redefining interaction paradigms, input-output channels, and the sequencing of generative and evaluative steps. It is worth noting that this reconfiguration also raises critical questions about maintaining design flexibility and preserving opportunities for serendipitous discovery, which remain vital for creativity. Therefore, the fundamental goal is to avoid rigidity in computational design processes; instead, architects should leverage algorithms and AI to empower themselves with the ability to flexibly customize, coordinate, and even disrupt workflows.
Through this multidimensional shift, the collaborative relationship between architects and computers is being rewritten. In an era increasingly shaped by computational technologies and artificial intelligence, the role of the architect must be reconceived―not merely as a traditional “creator” or “coordinator,” nor as a passive adopter of tools, but as a technology user, algorithm developer, and organizer of human–machine collaboration simultaneously. The ability of architects to recognize and understand their evolving position, and to gradually master the shift from using tools to collaborating with technologies, may emerge as one of the most essential competencies in contemporary design practice (Fig. 6).
On a deeper level, these three dimensions should not be seen merely as adaptive responses to new tools or technologies, but as indicators of a fundamental transformation in the structural logic of design methodology in the digital and intelligent age. As
Donald Schön (1992) argued, design is not only a process of constructing an object, but also of continuously revising and reorganizing the pathways of thought that generate it. Especially today, as open-ended web-based environments and low-threshold development ecosystems become the norm, what architects engage with is no longer a fixed tool, but an entirely reconfigurable workflow infrastructure. Within this context, digital technologies function more like a construction material―one that shapes the structural logic of the design mechanism itself.
6 Conclusion
This study has revisited the evolving synergy between architects and computational systems through the lens of design tool development. Rather than centering on tool functionalities or comparative evaluations, it adopts a reflective and practice-based approach, analyzing the development of four web-based generative design tools. From these cases, the paper proposes a theoretical framework comprising three interrelated dimensions: reconstructing intuitive media for interaction, adopting algorithmic and structured design thinking, and orchestrating the design process as a programmable system. These dimensions are not confined to any specific tools or scales, but highlight a broader shift in how architectural knowledge and agency are configured within computational environments.
This framework aims to contribute to ongoing discussions on human–computer collaboration by emphasizing the architect’s evolving role―not only as a tool user or algorithm designer, but as an active configurator of design workflows and computational systems. It underscores the importance of engaging with the deeper logic of computation to reframe the conditions under which design unfolds. While the framework has been grounded in large-scale urban and architectural tools, its core concepts―such as interaction design, logic structuring, and process configurability―may also offer insights for smaller-scale applications, including interior design, retrofitting, or component-based systems.
Rather than offering a fixed methodology or universal model, this research frames design tool-making as a reflective practice―one that enables architects to reassert their roles and conceptual authorship. Future research may extend this reflection across different design domains and practical projects, deepening the investigation into how computational thinking and architectural thinking can be more meaningfully aligned. Reimagining such agency may prove critical to shaping more coherent, flexible, and meaningful design practices in increasingly computational design environments.
2095-2635/2026 Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd.
PDF
(12922KB)
