1 Introduction
The spatial configuration of traditional rural communities in China deviates from the rigid geometric logic of modern planning, instead exhibiting organically evolved forms shaped by long-term interactions with the natural environment. residential buildings are typically organized through a sequential spatial hierarchy of courtyard, transitional space, and main dwelling area, forming an integrated courtyard typology that reflects and supports regionally specific ways of life (
Lin et al., 2024). In contrast, contemporary rural community planning is often dictated by top-down construction and development policies, which commonly mandate building on undeveloped land or require the demolition of existing self-built dwellings prior to the implementation of standardized, unified planning schemes. This design logic of “meeting the standards first, then fixing the form” often leads to overly standardized community layouts, resulting in pronounced homogenization and a marked loss of local identity (
Peng, 2015).
In response to this challenge, computational design methods have shown significant promise for application in rural planning contexts. As an integrated approach that combines rule-based logic, process simulation, optimization algorithms, and interactive data processing (
Schnabel, 2012), recent research demonstrates its capacity to support innovative strategies for optimizing spatial density, encoding vernacular spatial knowledge, and managing complex planning frameworks (
Er-Retby et al., 2025;
Kuang et al., 2025;
Wu et al., 2025). These developments offer new directions for the spatial organization of rural communities and the design of residential buildings. Notable examples include Li’s development of the Settlements Generating System for environment-responsive and self-adaptive rural settlement planning (
Li, 2021), as well as Zhang and Nauata’s application of the Wave Function Collapse algorithm (
Zhang et al., 2025) and the House-GAN framework (
Nauata et al., 2021), which introduce intelligent generative tools for residential layout automation. These innovative practices not only enhance design efficiency through algorithmic generation but also exemplify the trend of technological convergence in contemporary design methodologies. Nevertheless, the application of computational design to compact rural community planning remains constrained by several limitations. First, conventional heuristic algorithms struggle to efficiently manage the complexity of planning tasks that involve numerous spatial components and intricate topological relationships, often leading to overly complex generative processes. Second, the scarcity of region-specific rural design data hinders the development of learning-based models capable of emulating human design strategies. Current algorithms remain limited in their capacity to adapt to diverse geographic terrains and culturally specific spatial logics. Therefore, establishing a computational framework that not only addresses algorithmic inefficiencies but also integrates localized spatial intelligence into the design decision-making process represents a critical breakthrough. Such a framework could enable the transition from experience-driven to computation-driven planning while balancing the demands of standardization with the expression of spatial and cultural diversity.
Against this research backdrop, the design studio at Shandong Jianzhu University has spent a decade (2015–2024) developing the Unconventional Village model. Operating at two scales―community layout planning and individual residential building design―the model is grounded in extensive analytical and generative studies of traditional rural communities in China, with a specific focus on the Shandong region. For example, at the planning level, a rule-based approach was used to explore strategies for generating mass rural housing across the North China Plain (
Wang et al., 2023), which provides a research basis for the planning and design methods of plain areas in this paper; at the unit level, a hybrid method combining shape grammar with residential space requirements was used to produce customizable housing typologies (
Wang et al., 2022). Furthermore, an integration of shape grammar and Building Information Modeling was applied to generate modular mass housing designs in Shanli Chenjia Village, Zhaoyuan, Shandong (
Wang et al., 2025). The generative approach to unit-level layout from the latter project constitutes an important part of the residential layout of this study. As a computational design framework, the Unconventional Village model utilizes generative methods such as shape grammar to encode traditional rural spatial forms into formal geometric rules. These rules are intended to extract and formalize the logic underlying autonomous spatial evolution in traditional villages, facilitating its application in contemporary rural revitalization efforts. The model enables the scalable generation of planning schemes that reconcile standardization with spatial and cultural diversity. This paper will elaborate on the structure and application of the Unconventional Village computational model.
2 Literature review
2.1 An open design framework integrating community planning and residential building
The open design paradigm in community architecture emphasizes spatial openness and multi-level coordination. At its core, it constructs a hierarchical and computational system of design rules, enabling an organic integration between the spatial structure of communities and the form of individual housing units. This paradigm has evolved theoretically from an emphasis on “functional zoning” toward “holistic coordination.” Early research primarily focused on spatial openness at the planning scale (
Jacobs, 1961) and the mixing of functions (
Christopher, 1965), whereas contemporary studies highlight the establishment of multi-level regulatory frameworks linking planning and architectural design (
Talen and Lee, 2018). This shift is driven by the limitations observed in traditional community planning’s rule systems. Earlier planning approaches were characterized by excessively rigid regulations that over-standardized building forms and technical requirements, thereby restricting the flexibility and creativity of both designers and residents during implementation (
Foster, 2018). Moreover, local cultural identities were insufficiently translated into actionable design parameters (
Talen, 2012), resulting in a disconnect between newly constructed communities and their original contexts (
Turner, 1976). Concurrently, an imbalance between centralized control and individual autonomy has led to a lack of systematic integration across overall community layouts, group organizations, and individual housing designs. At the housing unit level, standardized dwellings often fail to accommodate diverse user needs (
Duarte, 2001), compounded by the absence of flexible mechanisms for self-construction or renovation (
Cozzolino et al., 2017). To address these challenges, contemporary research has developed three representative design paradigms, each embodying distinct logics of rule construction and adaptability.
The morphological coding approach establishes a hierarchical system of rules that progresses from the macro-level layout to the micro-level form, ensuring coordination between the overall community structure and individual building typologies. A key technique involves typological derivation (
Ðokić, 2009), where morphological rules―such as street cross-sections, building setbacks, and facade proportions―are extracted from local architectural prototypes and converted into computational design parameters (
Ben-Joseph, 2004;
Katz, 1994;
Veras and Amorim, 2005). In the representative New Urbanism project in the United States, Seaside, this approach not only preserves considerable design flexibility but also significantly enhances spatial legibility (
Dancy, 2007;
Macedo, 2016). The main advantage lies in the clarity and ease of implementing the rules; however, care must be taken to avoid morphological rigidity resulting from excessive standardization.
The flexible framework-infill model employs a collaborative frame plus infill strategy, where a fixed framework is established through infrastructure networks such as streets and utility systems to define the basic spatial order (
Domińczak, 2021). At the same time, flexible spaces are reserved at the residential building level to facilitate residents’ autonomous adaptation and self-organization (
Al-Hindi and Staddon, 1997; Marshall, 2011). The rule hierarchy in this approach generally consists of three levels: Level 1 rules govern the hierarchy of public spaces, Level 2 rules regulate the modular combinations of building units, and Level 3 rules allow for variations within individual units. In projects such as the Quinta da Malagueira social housing by Portuguese architect Álvaro Siza (
Mota, 2015) and the Aranya low-cost community by Indian architect Balkrishna Doshi (
Shekhawat, 2019), this model has demonstrated the capacity to generate spatial variations that are five to ten times greater than the basic typologies. It is particularly suitable for contexts that require the preservation of historical urban fabric or the accommodation of phased development (
Duarte, 2001).
The meta-rule self-organizing model is the most experimental approach, centered on transforming traditional planning elements into a modular “rule menu” that residents can select from and adjust, such as land use and construction standards, to develop personalized designs (
Gowdy, 2009;
Singh, 2010;
Talen, 2009). Key features of this model include the flexibility in rule selection―for example, parametric options related to public space ratios and material recycling rates―the iterative application of rules, where outcomes are continuously fed back into the system for refinement, and a mechanism for social collaboration that fosters neighborhood interaction through shared facility governance. The model’s implementation in the Oosterwold project in the Netherlands has demonstrated a level of spatial diversity significantly greater than that found in traditional communities. However, it also requires residents to exercise high levels of autonomy and engage in collaborative management (
Cozzolino et al., 2017).
Although these three design strategies differ, they all share several fundamental principles. First, there is a hierarchical interrelation that establishes a seamless control logic linking community, group, and residential buildings, thereby preventing disjunctions between planning and architectural design. Second, the rules are made computable by converting parameters to enable precise mapping between planning criteria and building forms. Third, openness in implementation is ensured by providing interfaces for rule adjustment at key hierarchical levels, which helps strike a balance between professional control and public participation. These explorations not only offer a robust methodological foundation for addressing the tension between standardized planning and regional contextual adaptation but also form the theoretical basis for developing the “Unconventional Village” design model presented in this study.
2.2 Computational design methods in building layout
Computational design methods in architectural planning and layout can be broadly categorized into three technical frameworks (
Caetano et al., 2020;
Hua et al., 2024). The first is parametric design, which enables iterative updates of form through variable manipulation and demonstrates clear advantages in handling complex terrains (
Janssen and Stouffs, 2015). The second is generative design, encompassing rule-based systems (
Caetano et al., 2020)―including but not limited to shape grammar, L-systems, cellular automata, and multi-agent systems―as well as optimization techniques and metaheuristic algorithms such as genetic and evolutionary algorithms (
Osman and Kelly, 1996). The third is a machine learning-driven paradigm (
Du et al., 2020), primarily involving neural network architectures like convolutional neural networks, generative adversarial networks, and Transformer models. These are widely applied in various domains, including cultural heritage preservation, performance optimization, and complex decision-making tasks. Data-driven design typically facilitates AI-assisted workflows by learning from relevant datasets (
Wu et al., 2019). Among these, layout generation methods frequently rely on the RPLAN dataset, which contains approximately 80,000 real residential floor plans with detailed annotations (
Wu et al., 2019). Hua, leveraging extensive apartment layout data, explored multiple possibilities for adapting floor plans to permanent structural frameworks and employed artificial intelligence to generate conceptual architectural designs (
Hua et al., 2024).
In the field of architectural planning and layout, researchers have progressively developed multiple computational design approaches to enhance the efficiency and adaptability of spatial configuration.
Zhang et al., 2021 proposed an approach incorporating regional adaptability and outdoor thermal performance to optimize the courtyard layout of rural houses in the gully regions of China’s
Halatsch et al., 2008 introduced an urban design method based on shape grammar, constructing a hierarchical set of rules to achieve design control across macro-to-micro scales, thus offering theoretical foundations for digitizing the spatial logic of traditional villages.
Wang et al., 2020 combined shape grammar with urban morphological analysis to propose a rule-based generative framework integrating top-down and bottom-up approaches, utilizing the CityEngine platform to rapidly generate three-dimensional scenes that conform to typical block morphologies in Nanjing. Meanwhile, metaheuristic algorithms have increasingly become essential techniques in this domain. Existing studies predominantly employ methods such as genetic algorithms, simulated annealing, particle swarm optimization, and ant colony optimization (
Sari and Jabi, 2024). These approaches facilitate the automated evolution of design schemes by encoding floor plans into computationally interpretable functions. For instance,
Liu et al., 2025 employed conditional generative adversarial networks to generate complete village layouts based on traditional prototypes and enhanced generation quality by tuning parameters and network structures. However, this method primarily evaluates output quality based on Pix2Pix-generated images, without adequately considering the design consistency between generated layouts and architectural planning principles. Additionally,
Wang et al., 2017 demonstrated that combining shape grammar with traditional village housing layout generation yields significant results. Nonetheless, as constraints become more complex, traditional global optimization methods face the challenge of a combinatorial explosion in the search space (
Zawidzki and Szklarski, 2020). To overcome this,
Peng et al., 2023 proposed a cost-aware combinatorial optimization approach for urban planning by integrating genetic algorithms with random forests and principal component analysis. More recently,
Zhang et al., 2025 developed an automated program implemented with Rhino and Grasshopper based on the wave function collapse (WFC) algorithm―a constraint-solving technique inspired by quantum mechanics―to generate and evaluate rural housing layouts. This method systematically analyzes geometric and topological constraints, effectively accommodating diverse residential requirements. Despite these advances offering novel solutions for floor plan optimization under complex constraints, algorithm-dependent design methods frequently neglect specific user design contexts (
Jiang et al., 2024). Consequently, they often exhibit suboptimal performance and low efficiency across varied urban environments and struggle to adapt to complex terrains and cultural contexts, thereby complicating the design generation process.
In the domain of path planning research, early efforts predominantly utilized bio-inspired optimization algorithms, including genetic algorithms (GA) and particle swarm optimization (PSO). GA is particularly effective for multi-objective optimization tasks. By emulating evolutionary processes, GA achieves global optimal solutions in adjusting building orientation and visual field configurations, balancing factors such as daylight access, landscape views, and evacuation distances (
Choi and Cho, 2010;
Xu et al., 2019). More recently,
Mao et al., 2024 proposed five advanced optimization algorithms specifically aimed at optimizing path layouts within residential spatial configurations. Dijkstra’s algorithm, a classical shortest path method, identifies the optimal route from a start node to an end node by traversing weighted graph nodes. It is frequently employed in residential planning for road network optimization, such as computing the shortest connection paths between neighborhood exits and housing units, and analyzing road network hierarchies through superimposed path flow analyses (
Huang et al., 2022). The wool thread model, developed at the Institute of Lightweight Structures and Conceptual Design (ILEK) in Stuttgart, serves as a generative form-finding approach for optimizing path networks (
Otto and Rasch 1995).
Lopes et al., 2014 pioneered the application of the wool thread model in generating urban layouts, street networks, and settlement patterns, demonstrating how organic complexity can inherently embody optimization and logical order. This model challenges conventional conceptions of orderly, high-performance, and optimized urban forms, presenting a notable contrast to the rigid modernist grid system.
In the domain of architectural floor plan optimization, current research primarily addresses two fundamental categories of constraints: topological and data-driven. Concerning topological constraints,
Rodrigues et al., 2013 and
Nisztuk (2019) employed evolutionary algorithms to ensure the topological validity of design proposals, while
Hua (2016) combined statistical region merging, subgraph matching, and simulated annealing techniques to simultaneously satisfy both topological and geometric constraints.
Shi et al., 2020 concentrated on resolving spatial adjacency constraints, whereas
Laignel et al., 2021 translated functional requirements into constraint conditions and utilized genetic algorithms to optimize critical parameters such as room lighting, symmetry, and circulation organization. Regarding data-driven approaches, there is a clear progression from traditional algorithms toward deep learning techniques.
Wu et al., 2019 implemented data-driven methods for the intelligent generation of apartment floor plans, and
Hua et al., 2024 developed a comprehensive database of floor plan layouts adaptable to various structural types based on extensive apartment layout data. With the advancement of artificial intelligence, the automatic generation of architectural floor plans has seen significant improvements. Building on the generative capabilities introduced by
Isola et al., 2017 through the Pix2Pix model,
Huang and Zheng (2018) pioneered the application of the pix2pixHD model for architectural layout recognition and generation, marking a major breakthrough in the application of machine learning within architectural design. Since then, researchers have developed a range of datasets and training strategies to support the use of Conditional Generative Adversarial Network (CGAN) in architectural planning, floor plan design, and façade generation. For instance,
Uzun et al., 2020 trained a DCGAN using Palladian Plans and evaluated its performance. The Graph2Plan model, introduced by
Hu et al., 2020, was the first to incorporate graph neural networks into the generation of single-unit apartment layouts, allowing for multiple design constraints including user-defined boundaries, functional zones, and topological relationships.
Rahbar et al., 2022 advanced this line of research by integrating rule-based and data-driven methods through the use of a quadtree algorithm to achieve intelligent spatial segmentation and hierarchical organization. More recently,
ÖZMAN and SELÇUK (2023) employed HouseGAN and related generative models to produce large-scale housing layouts.
Carrera et al., 2024 trained three most advanced existing models for generating interior plans using PUBLICPLAN. These methods enable deep learning models to effectively extract visual features from training data, learn visual similarity patterns, and rapidly generate floor plans that closely align with the statistical properties of the original datasets (
Park et al., 2024).
However, CGAN still faces many challenges in practical application. These include limitations in data accuracy, restricted accessibility of supporting tools, uncertainty regarding the real-world applicability of generated schemes, and inadequate responsiveness to non-standard design conditions (
Lin et al., 2023;
Min et al., 2023). For instance, in the research on the generation framework of three-dimensional architectural forms developed by Zhou, CycleGAN produced fragmented or irregular building contours when simulating existing architectural features, reflecting deficiencies in edge detection and geometric learning (
Zhou et al., 2023). In the domain of floor plan generation, Aalaei overcomes the traditional “rasterization-post-processing” method and vectorizes the output based on pixels, but it faces great challenges when dealing with discrete spatial representations, especially it becomes difficult to find the global optimal solution (
Aalaei et al., 2023). Technically, CGAN commonly use rasterized floor plans as input, resulting in relatively homogeneous datasets (
Uzun et al., 2020) that rely on structured features such as category labels, external building contours, and functional zoning layouts (
Parente et al., 2023). While such conditionality enhances output specificity, it tends to destabilize training when dealing with diverse residential typologies (
Aalaei et al., 2023). Furthermore, due to the scarcity of irregular boundaries and heterogeneous samples in real-world datasets, CGAN models often lack generalizability in unconventional spatial scenarios (
Park et al., 2024), limiting their ability to address personalized spatial requirements and complex environmental constraints. In contrast, despite their relatively low generative efficiency, rule-based expert systems allow for explicit definition of building components―such as walls, doors, windows, and room types―and are better suited for tailoring outputs to specific needs without depending on extensive training datasets. This improves the practical usability of generated solutions and facilitates their real-world implementation. Such an approach is particularly significant in addressing the complex and highly heterogeneous spatial needs of rural communities in China, while also supporting a multidimensional balance among form, function, and cultural continuity. Therefore, to meet the evolving demands of rural community planning in China, there is a pressing need for the development of efficient, user-oriented design tools that can preserve local cultural characteristics and support spatial generation across diverse topographies and cultural contexts.
3 Methods
This paper aims to develop a systematic computational design model for compact rural communities in China, termed the “Unconventional Village.” By adjusting the collaborative mechanism between professional control and public participation, establishing a parametric mapping from planning indicators to architectural forms, and implementing a hierarchical control logic spanning community, group, and residence levels, the model rapidly generates diverse design schemes that balance compact land use with cultural continuity. This provides a technical foundation to address the fundamental conflict between standardized planning and local contextual adaptation. In constructing the “Unconventional Village” model, the research adopts a dual approach combining case-based and rule-based methodologies. The case-based approach offers theoretical and empirical foundations for the model’s constitutive logic and structural framework, serving as the basis for the overall design. Meanwhile, the rule-based approach plays a central role at the operational level by abstracting and refining context-specific rule systems to guide spatial generation and rule formulation, thereby ensuring the feasibility and adaptability of planning schemes.
3.1 Case-based method
This study employs a multi-scale case analysis approach to systematically examine several representative examples of community planning and residential design, including the Seaside community in the United States, the Aranya low-cost housing community in India, the Quinta da Malagueira community in Portugal, and the Oosterwold project in the Netherlands. Figure 1 shows the general layout of each case. The building control regulations associated with these projects are discussed in the works of
Dancy (2007),
Mota (2015),
Shekhawat (2019), and
Cozzolino et al., 2017. These cases provide critical references for developing a morphological framework tailored to compact rural community planning in local contexts. The framework comprises two main components. The first focuses on generating community planning layouts by establishing a hierarchical spatial grammar system that links three levels of control―streets, groups, and buildings―and by designing combinatorial computational rules for infill modules to produce diverse aggregate morphologies. The second component addresses user needs by generating customized, detailed layouts of residential buildings, ensuring the open implementation of self-built housing, with particular emphasis on the design of functional configurations and spatial relationships.
3.2 Rule-based method
Shape Grammar is a rule-based modeling approach extensively applied in urban and architectural design. Its core concept involves iteratively generating complex geometric structures from an initial shape by defining a set of production rules (
Stiny, 1980). Originally developed by George
Stiny (1971) for spatial computation and visual processing, it has since evolved into an effective tool supporting complex design generation. By systematically formulating compositional rules governing form combinations, such as block subdivision and building layout arrangements, Shape Grammar facilitates flexible urban design solutions. For instance, Duarte proposed a four-stage method involving territorial reading, basic geometric shapes, unit definition, and material attributes to generate urban designs (
Beirão and Duarte, 2005). Additionally, Shape Grammar is widely utilized in generative architectural design, where existing forms are iteratively transformed through addition, subdivision, or grid-based strategies to produce new building typologies (
Colakoglu, 2005), exemplified by the mass customization of housing in Siza’s projects (
Duarte, 2005). The rule system within Shape Grammar can be understood as a set of generic design operations; designers can extend this rule set by incorporating new customizable grammar rules and achieve precise adaptation to specific environmental contexts by restricting rule parameter ranges. This rule-based spatial organization framework constitutes a critical methodological foundation for the development of the “Unconventional Village” model in this study.
3.3 The Unconventional Village model
The model comprises two main subsections: community layout planning and residential building layout design. The first subsection, community layout planning, is composed of two elements: road networks and group form. The second subsection, residential building layout design, includes spatial combination constraints and spatial transformation operations.
3.3.1 Community layout planning subsection
The community layout planning subsection is built upon a rule-based planning system designed to address complex decisions related to road network organization, terrain adaptation, and cultural interpretation, enabling the generation of community morphologies that harmonize with the village’s character. This module first takes into account environmental elements and site topography, integrating various special constraints to establish a multi-level rule framework. The design process follows a procedural workflow from street to group to building, and applies to three primary scenarios: (1) flat terrain with relatively regular site boundaries, representing the most common case in practice; (2) mountainous terrain, where elevation differences must be addressed; and (3) special constraint conditions such as waterfront landscape considerations, integration of old and new buildings, mixing of multi-story and low-rise structures, and the blending of residential and productive functions. Using the site road network as input, the module determines the shape, orientation, size, and location of groups through road network analysis. Depending on the design scenario, group-scale partitioning is performed to identify appropriate building typologies for each group. This process constructs a preliminary spatial grammar system with hierarchical control indicators at the street, group, and building levels. Building samples used for group infill serve as design templates, which, guided by shape grammar rules, undergo spatial translation and combination operations to generate multiple alternative community morphology schemes. The shape grammar-based design rule system assigns attributes such as dimension, proportion, and function to each sample, satisfying diverse villagers needs and facilitating retrieval and recommendation within the design system. The specific operational steps of this module are detailed as follows.
(1) Streets and groups. The methods for dividing streets and groups vary depending on the terrain. For flat terrain, the shape, orientation, size, and location of blocks and groups are typically determined by the road network. Path connections, as a crucial element of the pedestrian circulation system, are designed based on shortest path principles to maximize user convenience. The process begins with an analysis of external environmental constraints, including site resources, surrounding roads, and functional distributions, selecting key resource points as the foundational framework for the road network layout. Subsequently, shortest paths are computed between multiple selected points within the site. To improve computational efficiency, the Wool Algorithm―a heuristic method for pairwise point connection―is employed, reducing redundant calculations by fitting closely located paths. Figure 2 presents a complete example of the Wool Algorithm, demonstrating its capabilities in optimal pathfinding and site partitioning. The white space in the original site represents water bodies and open areas, while marked points denote public space nodes extracted through site analysis. These nodes are modeled in Grasshopper using the point tool and linked to reference points. Iterative computations are then performed via the Wool Algorithm, with parameters such as point attraction, line density, and line tension adjusted within the group component, ultimately generating the road network and corresponding group divisions as shown. For mountainous terrain, to balance efficient terrain use and economic considerations, this study applies a triangulated irregular network (TIN) method based on digital terrain models (DTM) to divide terraces and optimize earthwork balance. The approach aims to minimize earthmoving volumes by abstracting cut and fill areas into a supply-and-demand network, with a solver generating the optimal allocation scheme. Constraints include balancing cut and fill volumes, adhering to slope safety limits (≤10%), avoiding ecologically sensitive zones, and meeting soil suitability requirements, with necessary manual adjustments and corrections incorporated.
(2) Group form. Housing, serving as the infill module within groups, directly influences the formation of collective morphologies through its configuration. The planning layout is achieved by establishing a hierarchical system of generative design rules: primary rules govern the relationship between building arrangements and group boundaries; secondary rules coordinate the integration of infill modules with public spaces and the road network; tertiary rules regulate building density as well as connectivity within and beyond the community. These rules can be dynamically adapted to various site contexts and are embodied in the “Unconventional Village” design series (Fig. 3). This series includes a 2016 field investigation of Seaside Town in the United States, in order to explore the open community planning model in depth, so the research on scheme design was interrupted that year. Specifically, the application of the “Unconventional Village” framework can be categorized into the following three scenarios.
In flat terrain planning, the road system generates largescale groups based on quantitative analyses, including pedestrian flow statistics, average travel distances, and existing village parameters. These groups exhibit symmetrically balanced building infill patterns, with spatial organization governed by hierarchical generative rules: the primary unit comprises four modules arranged around a cruciform pedestrian pathway, with the courtyard at the upper-left corner serving as the origin point and modules placed clockwise to form a four-quadrant layout; the secondary unit is formed by combining four primary units according to the same rules; and the tertiary unit further integrates four secondary units, thereby creating a clearly hierarchical planning framework (Fig. 4(a)).
For planning layouts on mountainous terrain, the generation rules require differentiated treatment. In the case of terraced mountainous terrain, small rectangular groups formed by road divisions follow a hierarchical three-tier rule system: primary rules govern the number of short-edge modules filling the group and their arrangement in relation to the terrain; secondary rules combine two primary modules side by side to coordinate with adjacent public spaces and roads; tertiary rules further integrate secondary modules to form composite units. For irregular terrain, an adaptive grid is first established along the plot boundary. Multiple courtyards are combined following the hierarchical three-tier module filling rules described above to ensure the organic integration of form and terrain (Fig. 4(b)).
In the context of rural planning under special constraints, it is necessary to first redefine the constituent elements of infill modules and establish an adaptive system of primary rules (Fig. 5). Spatial configurations are then derived in accordance with the secondary and tertiary rule logics tailored to the topography of the plains: (1) In production-living integrated areas, the infill modules comprise a variety of functional structures such as residential units and warehouses. The primary rule adopts a hybrid “residential + productive” module that integrates small-scale production spaces within traditional courtyard layouts, resulting in mixed-use spatial clusters. (2) In waterfront areas, platform elements are introduced into the infill modules, and primary rules employ a dislocation arrangement and elevated platform designs to enhance visual and spatial permeability toward the landscape. (3) In the mixed areas of multi-story and low-rise buildings, the infill module is defined as a combination of three-story structures and flat-roofed single-story units, with the flat roofs functioning as elevated public spaces to mediate scale differences. The primary rule defines differentiated combinations of multi-platform modules. (4) In zones where new and old developments coexist, infill modules are adaptively redesigned based on traditional courtyard prototypes. The primary rule ensures spatial coherence by incorporating morphological transition nodes and extending the existing road network.
During the sample selection process, a hierarchical optimization mechanism is established to ensure the feasibility of the planning: at the first level, each module must include at least one pedestrian pathway connecting to other homesteads to prevent isolated units; at the second level, the layout of roadside modules is optimized by dynamically adjusting rules, including extension and shortening strategies―specifically by adding or removing one standard module along the vehicular road direction―thereby effectively minimizing dead-end roads; at the third level, utilizing the secondary sample database, the number of infill modules is appropriately adjusted to reserve adequate space for various tiers of public open spaces (such as green areas and planting zones), thereby optimizing the allocation of community public functions.
3.3.2 Residential building layout design subsection
The building layout subsection generates diverse floor plans within a predefined architectural form framework according to user requirements. It employs shape grammar rules to establish a functional space adjacency network and applies dimensional constraints to regulate spatial scale relationships. This approach effectively manages the adjacency constraints of functional spaces and the spatial manipulation operations involved in floor plan generation. Functional space adjacency relationships restrict local neighborhoods during layout generation, while the outcomes of spatial manipulations provide a reference framework for shape grammar–based floor plan design. The specific operational steps of this module are as follows.
(1) A prioritized hierarchical rule system is established to govern functional space combination constraints. The primary requirements include having at least one south-facing bedroom, prioritizing the placement of living rooms on the south side, and positioning kitchens with direct external windows. Secondary rules address hygiene and safety considerations, such as locating bathrooms away from dining and kitchen areas and orienting staircases to face north (
Li, 2024). Concurrently, standardized dimensional modules are employed to harmonize the proportions of functional spaces, ensuring the feasibility of modular construction.
(2) The moving operations for floor plan generation are primarily based on analyzing spatial relationship combinations using design corpora and user-specific customization requirements. The spatial configuration is defined based on the principle of exhaustive enumeration. Figure 6 shows the design and implementation of a function combination generation program based on recursive algorithm. The program encapsulates the core operations by defining a class, in which a core class manages essential operations such as element initialization, basic data manipulation, and formatted output, thereby establishing a foundational framework for organizing spatial layout data. Each functional space―such as bedroom, dining room, living room, bathroom, kitchen, and porch (labeled as 1, 2, 3, 4, 5, and 6 respectively in Fig. 7(a))―is assigned a distinct numeric code and visually represented with different colors. Building on this structure, the program implements a recursive algorithm that explores all potential combinations of spatial codes through a depth-first search strategy. Valid combinations that meet predefined constraints are stored in a global list. The final output is generated via a computational module invoked by the main function, thereby enabling a complete index enumeration of spatial layout arrangement. The above-mentioned spatial relationship combinations reasoning process mainly completes the formulation of relevant rules with the help of the logical structure of tree diagram, and Fig. 7(b) shows the reasoning process. First, the architectural plan was discretized into a basic grid system. Within this system, the positions of the residential unit, courtyard, and the entrances to both were identified. Next, vertical circulation elements and functionally constrained spaces―such as south-facing bedrooms and living areas―were defined. Based on spatial adjacency relationships, the functions of surrounding spaces were inferred, resulting in a complete functional layout plan that incorporates the residence, courtyard, and auxiliary rooms.
4 The community planning of new mass housing in Shanli Chenjia Village
Facing a severe population aging issue and the recent rise of rural tourism, local developers in 2023 planned to construct a series of multifunctional mass housing units in Shanli Chenjia Village, Zhaoyuan, Yantai, designed to accommodate both elderly villagers and tourists. The site, oriented east-west, features an elevation difference exceeding 20 m (Fig. 8). According to local planning regulations, new buildings are required to be at least three stories tall, with a maximum floor area of 60 m2 per unit. Against this backdrop, the studio employed the “Unconventional Village” model and conducted experimental research using two distinct computational design approaches to explore generation strategies for community planning and residential building layout design, ultimately yielding a comprehensive planning and design scheme.
4.1 The results of community planning and layout design
Experiment 1 follows the regulation for accessible ramps, which requires a minimum ramp length of 15 m for every 1 m of elevation gain. Accordingly, in Grasshopper, a circle (Fig. 9(a)) with a 15-m radius is drawn using the Circle component, centered at the midpoint of the site entrance. This circle intersects with the contour line at the next 1-m elevation increment, and the intersection point is designated as the center of the subsequent circle. Using the curveIntersection or shatter components, the centers of subsequent circles are sequentially determined. These centers are then successively connected using the Line component to form a predefined ramp path (Fig. 9(b)). Next, Weaverbird’s subdivision function is applied to further subdivide and generate additional path nodes, ultimately creating a complete subdivided mesh across the entire site (Fig. 9(c)). By connecting points at identical elevations, horizontally continuous mesh lines are generated (Fig. 9(d)), with each line segment serving as an accessible path. Starting from the site entrance, the ListItem component extends the path bidirectionally along the mesh lines, with a step size ranging from 1 to 3 segments, until reaching the site boundary. Finally, integrating the Region component with the dimensional rules for group divisions completes the overall road network and spatial partitioning of the groups (Fig. 9(e)).
Secondly, the areas delineated by the road network are designated as reserved group parcels, within which adaptive boundary layout rules are applied to guide building infill design. Specifically, for each plot, two corner points forming angles closest to 90°, or plot edges that are approximately parallel, are identified as reference baselines. Based on these reference directions, the largest possible rectangular grid fitting within the plot is generated. After excluding the area occupied by this rectangle, the remaining portion of the plot typically forms an approximate triangular shape. Subsequently, the longest edge of this triangle―or alternatively, the edge closest to the main road―is used as a directional reference to construct a second rectangular grid orthogonal to the first, which fills the residual space (Fig. 10). Through this sequential generation of two maximal rectangular grids, the approach ensures efficient utilization of irregularly shaped plots and establishes a structured framework for the orderly arrangement of individual buildings.
The infill modules use the grid generated in the previous steps as a reference framework, applying shape grammar to deconstruct the traditional courtyard house typology into fundamental unit combinations oriented toward the four cardinal directions: east, south, west, and north (denoted as E, S, W, and N, respectively). Each unit can occupy various positional states within the grid, such as interior/exterior and upper/middle/lower levels, and is categorized into six types (Fig. 11(a)), constituting the primary dataset. Based on this, three or four building units of different orientations are selected from the primary dataset and combined to generate infill modules that embody the spatial characteristics of traditional courtyards, thereby producing layouts that harmonize with the local courtyard patterns. To optimize southern daylight access, northern units are designated as mandatory components, consistently adopting a three-story pitched roof form. Six typical northern unit configurations (N6) serve as the foundational benchmarks. Building on these, feasible combination rules for units in other directions, occupying different spatial positions, are developed to form a secondary dataset (Fig. 11(b)), encompassing 24 distinct combination patterns. The heights of units in directions other than north may vary from one to three stories, with roof forms either flat or pitched (Fig. 11(c)). To enhance spatial interaction within the courtyard, an additional rule is introduced: if no platform oriented toward the courtyard exists within a combined module, a platform is added to the south façade of the northern building unit (Fig. 11(d)) to create an interaction space. Ultimately, the system selects combination modules from the secondary dataset that maximize spatial occupancy within the grid to perform infill, enabling automatic generation of the overall floor plan layout.
Experiment 2 aims to optimize both terrain utilization efficiency and cost-effectiveness by employing the irregular triangular network method based on a digital terrain model for earthwork balance optimization. The procedure is as follows: Elevation point data in *.dat format is first imported into the South CASS software, where it is processed and visualized within the drawing module. An irregular triangular network is then generated from the elevation points (Fig. 12), and design platforms for each terrace level are defined using engineering drawing tools in accordance with target elevations. Based on calculations of cut and fill volumes, the positions and heights of these platforms are iteratively adjusted to achieve dynamic earthwork balance, ultimately dividing the site into several leveled terraces (Fig. 12) while simultaneously generating a basic road network.
At the group design scale, when the elevation difference between adjacent groups reaches 3 m, terrace-style building foundations aligned with the terrain are employed, expanding group widths beyond 12 ms to accommodate stepped residential layouts. When the building floor height matches the terrace height, the roof of the lower terrace buildings aligns flush with the ground of the upper terrace, creating continuous circulation paths between ground and roof levels. This spatial configuration facilitates the construction of a vertical traffic evacuation system adapted to varying elevation differences, enhancing the efficiency and compactness of vertical building organization. Accordingly, buildings are staggered between terraces to form effective traffic connection spaces. Within the experimental terrace limits, two spatial relationships between buildings on adjacent terraces―“separated” and “adjacent”―are defined, while configurations where the upper story of a lower terrace building cantilevers closely over another building are excluded. Ultimately, 28 feasible terraced staggered combination patterns were generated (Fig. 13), providing design references for group layouts under diverse terrain conditions.
Building upon the preceding analysis, this experiment further develops a three-tiered building combination library tailored to terrain constraint conditions for generating spatial layout typologies at the group scale. Initially, a basic grid system is established considering a plot width of approximately 12 m, a standard building footprint of 6 m × 9 m, and a mandatory minimum fire separation distance of 6 m between buildings. In this grid system, the right edge corresponds to the northern orientation of the actual site. Each building unit is confined within a 4 × 5 grid, where each grid cell measures 3 m, subject to constraints arising from its adjacency to terrain boundaries and corridors. Under these parameters, building units can be translated into nine distinct positions within the grid, with each position further generating eight spatial configurations based on the positional relationship between the corridors and the building unit (such as attachment to a single side or placement between two adjacent edges), thereby constituting the first-level combination library (Fig. 14(a)). Building upon this, two building units are horizontally joined to form the second-level library, governed by the following splicing rules: (1) maintaining a minimum fire separation distance of 6 m between adjacent units; (2) prohibiting complete overlap of short edges when units are closely arranged to enhance the variation in roof forms; (3) removing redundant mirror-image duplicates; and (4) excluding combinations featuring more than three corridors to mitigate interference of circulation with building functions. Figure 14(b) shows the results of partial combination. The third-level library is generated by combining pairs of second-level samples (Fig. 14(c)). To reduce the number of design samples and improve overall spatial logic, the following rules are established for the group configurations generated from the combination of two third-level morphological typologies (Fig. 14(d)): (1) For buildings located on the uppermost and lowermost terraces, inter-building outdoor corridors should follow the nearest-neighbor principle and be placed only along the short sides or the courtyard-facing long sides. For buildings situated on the middle terrace, outdoor corridors are generally positioned on the short sides. When two such corridors are located within a distance of two grid units, an alternative vertical placement along the long side may be considered; (2) Continuity between the corridors on the top and bottom terraces should be preserved to the greatest extent possible; (3) When adjacent buildings on the middle terrace are aligned without staggering on their south-facing short sides, daylighting to the northern buildings may be obstructed. In such cases, the buildings causing solar obstruction should be removed from the configuration. Finally, a bounded set of feasible design configurations is derived through comprehensive solar analysis.
4.2 The results of residential buildings plan layout design
Experiment 1 transforms individual floor plans into planar forms exhibiting fourfold symmetry within a 3 × 3 grid system. The grid is based on a 1.8-m module, which meets the standard width for accessible circulation. Functional modules are progressively combined and transformed through exhaustive positional enumeration guided by shape grammar rules. Initially, four living area geometries and seven activity area geometries are imported separately via geometry inputs and merged using two Merge components into two lists: the Living Area List (List A) and the Activity Area List (List B), establishing the initial geometric modules. Subsequently, the Cross Reference component connects List A and List B to inputs A and B, respectively, with the mode set to Cartesian product, thereby generating all possible combinations. The Graft operation is then applied to List B to create independent branches for each activity area geometry, facilitating individual processing. Prior to the combination, the width of each living area geometry is computed using bounding box and box dimension functions to determine the precise translation distance needed for the activity area geometries. According to these calculations, each geometry in List B is translated to align accurately with its corresponding living area geometry. Following translation, the combination of Graft and Flatten operations produces one-to-one pairs in the Cross Reference output, ensuring correct pairing of geometries. Finally, the Join function within the Transform module concatenates these pairs into complete floor plan configurations (Fig. 15).
Experiment 2 employs the tree-based logical deduction capability of shape grammar to derive functional area layouts, establishing a mapping between shape rules and functional modules (Fig. 16(a)). Based on this rule set, corresponding families are created in the Revit platform, and the final floor plan design is realized by placing these modular families accordingly (Fig. 16(b)). The detailed steps of this process can be found in the mass modular housing design research completed by
Wang et al., 2025.
5 Disscusion
5.1 Evaluation of experimental results
Following the design process described above, two design schemes, as Fig. 17(a), were produced. The design team conducted a systematic evaluation of these schemes to verify their fulfillment of the intended objectives and to select the optimal option accordingly. The evaluation employed a quantitative assessment framework based on the methodology proposed by Mahmoud (
Mahmoud and Youssef, 2022), with specific weights assigned to each evaluation criterion as shown in Fig. 17(b). A questionnaire survey involving ten researchers was conducted to determine the relative importance of each criterion. Subsequently, 30 local villagers participated in a satisfaction survey assessing the performance of the two schemes across all criteria. Satisfaction was rated on a five-point Likert scale: “very satisfied,” “satisfied,” “average,” “dissatisfied,” and “very dissatisfied,” corresponding to scores of 5, 4, 3, 2, and 1, respectively. Scores ranging from 4 to 5 were considered “excellent,” while those between 3 and 4 were deemed “acceptable.” Finally, the fuzzy comprehensive evaluation method is used to evaluate the two schemes, and the results of experiment 1 and experiment 2 are 4.135 points and 4.521 points, respectively.
Based on the statistical results, experiment 2’s design received the highest score. To further investigate the relationship between the design outcomes and the original layout and units of the village, this paper compares some areas randomly selected in the village with the design result with the highest score. Spatial syntax analysis was employed to assess spatial integration, convex space connectivity, and axial diagrams (Fig. 18). The comparison reveals that the narrow spaces between closely situated houses in the existing village layout exhibit lower integration, while open areas demonstrate higher integration. This characteristic is reflected in the design scheme, where the central courtyard achieves a higher integration score, while the narrow spaces surrounding it score lower, effectively translating the original features. The axial diagram analysis indicates that the passage routes in the selected area consist of segments with varying degrees of connectivity, reflecting the winding nature of village roads and the historical arrangement of buildings. The central road in the design’s axial diagram shares this characteristic, highlighting its accessibility and multi-scale form. Regarding convex spaces, the number of divisions within the village’s convex spaces illustrates a fusion of spatial characteristics, with their shapes, sizes, and orientations exhibiting a natural and discrete quality. In the design, the convex spaces created by the combination of architectural units not only retain their natural characteristics but also have clearer delineations, suggesting an optimization of the design based on the continuation of village traditions. In summary, the best-selected design scheme has successfully achieved coordination and unity with the village, fulfilling the design objectives.
5.2 Standardized planning and local adaptability
The core challenge in rural community planning lies in reconciling standardized, replicable technical frameworks with the unique natural and cultural characteristics of local contexts. This study introduces a multi-layered rule-based system that establishes a dynamic equilibrium between technical rationality and vernacular intelligence, structured through a spatial grammar framework across the scales of streets, groups, and individual buildings. The integrated logic of this system operates on three hierarchical levels: First, at the community scale, computational analysis is used to generate a rigid spatial framework for road network planning. By incorporating terrain-responsive modeling techniques, the system enables a scientifically optimized distribution of street density that ensures transportation efficiency while preserving the spatial topology of traditional communities. Second, at the group level, architectural forms are generated through a parametric mechanism that translates the cultural logic of traditional courtyard configurations, such as proportional relationships and spatial hierarchies, into new built fabrics that resonate with the morphological character of the village. On this foundation, multi-tiered combinatorial rules define a flexible framework for the morphological evolution of collective forms. Within the rigid constraints of the road network and density indicators, diverse spatial configurations emerge through topological transformations of modular units, thereby reconciling the efficiency of standardized production with the adaptability required by regional specificity. Finally, at the scale of individual dwellings, a differentiated configuration system further mitigates the limitations of standardization. By offering a selection of alternative floor plan options, the model supports resident participation in housing layout decisions, enabling precise accommodation of heterogeneous household needs at the smallest unit of regulatory control. This synergistic mechanism―balancing rigid constraints with flexible infill―presents a practical paradigm for addressing the tension between standardized planning and the preservation of vernacular spatial culture.
5.3 Advantages and disadvantages of the Unconventional Village model
Leveraging the Unconventional Village model, the design studio has conducted a decade-long series of experiments on community design across diverse scenarios, including planning schemes for plains, mountainous terrains, and sites with special constraints. These outcomes offer exemplary design templates for rural community planning in China and exhibit notable advantages along three key aspects: Firstly, the Unconventional Village model operates without reliance on manually generated design samples as training data and does not require designers to have advanced programming skills, thereby avoiding complexity in the generative design process. This makes it particularly well-suited for generative design tasks lacking access to large-scale datasets. Secondly, rather than aiming to generate complex global plans, the model reframes challenging local constraint issues, such as terrain adaptation and spatial relationship optimization, into computational problems addressable through targeted design strategies. Simultaneously, the model embodies a systematic design thinking approach that balances the multidimensional factors of form, function, and culture. It satisfies hard constraints, including daylight and fire safety regulations, while preserving soft cultural elements like courtyard arrangements and regional architectural characteristics, substantially enhancing both the scientific rigor and contextual adaptability of the designs. Finally, to integrate with traditional workflows that rely on two-dimensional CAD drafting, the model provides interfaces for multi-tool integration and data exchange, enabling iterative data conversion and transfer. This facilitates more efficient collaboration and supports multi-criteria decision-making processes.
Conversely, with advancements in computational technologies such as data processing and machine learning, the paradigms and concepts of computational modeling have been continually evolving. Over the past decade, the studio has explored various strategies for automated design. For instance, it applied graph-theoretical methods to examine the spatial connectivity among public spaces, residential buildings, warehouses, and greenhouses in layout configurations. A 3 × 3 adjacency matrix was formulated under a set of constraint-based rules to structurally represent the spatial relationships between functional components (Fig. 19). In this model, the public space was designated as the central node to maximize network connectivity; each residential plot was required to maintain a direct edge with the public space, and warehouses were mandatorily connected to greenhouses. Additionally, the degree and corresponding weight of each node were assigned within defined ranges to accommodate functional variability. These matrix-based formulations grounded in graph theory provided the computational basis for the automated generation of grouped layouts using Python. However, bottlenecks have gradually surfaced regarding technical integration and interdisciplinary collaboration. A knowledge gap exists between algorithm developers and on-site designers, which hampers the ability of computational design tools to accurately address the complex, context-specific design requirements. Moreover, the pace of technological advancement often lags behind rural construction cycles, resulting in underutilization of digital tools and insufficient support for the practical implementation of projects.
6 Conclusion
The fundamental challenge in compact rural community planning in China lies in balancing standardization with regional adaptability. This study, grounded in case analyses of community planning and employing rule-based methods, proposes the Unconventional Village model. Culminating from a decade of experimental exploration, the “Unconventional Village” design series aims to effectively address the tension between regulatory control and local autonomy in rural communities characterized by complex environmental factors and diverse site conditions. The model considers various terrain types and special constraints while presenting practical design methods, demonstrating its applicability as a computational design approach in Chinese rural community planning.
(1) At the technical level, theories and methods for open design in Chinese rural communities remain underdeveloped, and fully algorithm-driven approaches are not yet widely accessible to designers in practice. There is a need for a decision-support model that can both overcome algorithmic efficiency limitations and incorporate the inheritance of regional knowledge and traditions. The Unconventional Village model features prototype-based integration and rule-driven generative computational design, integrating generative layout planning with multiple design evaluation algorithms within a design framework. This framework embeds regional spatial design rules by combining graph topology with computer-aided spatial analysis. Its multi-level rule system is compatible with multiple software toolchains and requires minimal data input, lowering technical barriers and providing a reusable methodological framework for rural design.
(2) At the planning level, introducing mechanisms for autonomous construction within planning control can effectively generate diverse morphologies, addressing the tension between standardization and local adaptability in Chinese rural community planning. The Unconventional Village model translates the spatial logic of traditional autonomous rural construction into computable parameters, establishing a precise mapping between planning metrics and building forms. It supports rapid iteration from site analysis to scheme generation, producing diverse alternative planning scenarios. This achieves distinctive spatial patterns that align with yet remain distinct from traditional rural forms, providing theoretical support for the digital practice of “New Vernacularism.”
(3) At the building unit level, residents exhibit varied demands for floor plans. Traditional single-type floor plan design models no longer satisfy contemporary social needs, necessitating diversified design schemes to accommodate users. The Unconventional Village model constrains the dimensions and adjacency of functional modules through rules while reserving interfaces for various combination patterns, enabling residents to select functional modules autonomously. This approach overcomes the lack of autonomy in individual building design, generating differentiated building layouts that maintain the overall integrity of community planning and respond to diverse household needs.
2095-2635/2025 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd.
PDF
(19057KB)
