基于深度学习的城市形态、街区形态与建筑群形态及布局生成：2018-2027 英文期刊文献综述

In response to the growing demand for intelligent solutions in urban planning, this study constructs a deep learning-based framework for generating intelligent urban morphology, effectively addressing pressing real-world challenges. At the outset, the study explores the core concepts of green and ecological principles within the evolution of contemporary urban forms, establishing a robust theoretical foundation for subsequent investigations. The study provides a detailed explanation of the practical application paradigms of deep learning, encompassing meticulously selected technical methodologies, carefully designed algorithmic structures, and an optimized parameter configuration system. Together, these elements form a comprehensive technological application framework. An innovative application of convolutional neural networks is introduced for the in-depth analysis and processing of urban street imagery. This advancement enables critical urban planning functions, including road network design, detailed analysis of building distributions, optimization of public facility layouts, and dynamic traffic flow analysis. These capabilities address the key limitations of traditional planning methods by enhancing intelligent analysis and precise decision-making. To evaluate the model's performance quantitatively, a systematic testing scheme is developed and implemented, covering various scenarios, including daytime and nighttime conditions. This approach ensures a comprehensive assessment of the precision and effectiveness of each functionality. The core significance and contributions of this study are encapsulated in its empirical findings. The proposed model achieves accuracy and fit metrics exceeding 93% across all testing dimensions, representing a significant advancement that provides robust and targeted support for urban planning practices. By integrating deep learning technologies into the intelligent urban morphology generation framework, the study successfully implements critical functions such as efficient road network planning and scientific analysis of building distributions. Furthermore, the study introduces cutting-edge technological tools and innovative methodologies to the urban planning discipline, advancing the development of intelligent urban planning. Its contributions are of profound value in both theoretical innovation and practical application, offering transformative potential for the field.

Amid China’s rapid urbanization, many historic districts face complex challenges, including fragmented traditional fabrics, disordered spatial morphology, and discontinuous street networks. To tackle these issues, this study proposes a multimodal deep learning framework that combines Generative Adversarial Networks (GANs) and Diffusion Models, establishing an integrated generation-optimization workflow for the renewal of historic districts. The methodology begins by using Pix2PixHD to generate high-precision fabric layouts, followed by fine-tuning a Diffusion Model through Low-Rank Adaptation (LoRA) to achieve diversified morphological expansion. The candidate proposals are quantitatively evaluated using a ten-indicator evaluation matrix that covers both architectural fabric and street network dimensions. Afterwards, these proposals undergo iterative optimization with a multi-objective framework to enhance both urban fabric morphology and network performance. The framework was validated through an empirical study of the Yuehe Historic District in Jiaxing. The results indicate that the generated schemes closely align with the original urban fabric. Compared with the existing expanded area (EA), the weighted comprehensive fitness score of the optimized scheme group improved from 0.66 to 0.89 ± 0.02 (a 34.8% increase), with the standard deviation decreasing from 0.07 to 0.02, indicating significantly enhanced stability. Deep learning balances morphological authenticity, generative diversity, and performance in historic district preservation and renewal.

Against the backdrop of rapid urbanization, urban form homogenization has become a prominent challenge, seriously eroding cities’ unique cultural identities and functional diversity. Traditional urban form design mainly relies on designers’ subjective experience, which often fails to strike a reasonable balance between the accurate representation of urban morphological attributes and the effective preservation of spatial diversity. To address this critical limitation, this study proposes a deep learning-based conditional generative control model, with metropolitan areas of developed cities around the world as the specific research context. To verify the effectiveness of the proposed model, we adopted a systematic research method: first, we built a 5-dimensional urban form evaluation framework (covering shape, scale, compactness, fragmentation, and proximity) based on landscape pattern indices; second, we constructed a high-quality urban morphology dataset; third, we verified the correlation between landscape pattern indices and generated urban form features; finally, we conducted comprehensive tests to evaluate the model’s accuracy and scalability. The experimental results show that in over 90% of the sample images, the Percentage of Like Adjacencies (PLADJ) index exhibited the strongest correlation with the urban morphological features generated by the model. Meanwhile, the proposed model achieved an average accuracy of 85.72% in generating urban morphological indicators across the five core dimensions, proving its good performance in simulating urban forms. This study reveals that although single-dimensional indicators are context-dependent, they are closely correlated with the overall characteristics of urban form. The research innovatively integrates landscape pattern quantification with conditional generative models, providing a practical technical tool for urban planning. Its limitations lie in the narrow research context, and there is still room for optimization in indicator selection and model efficiency.

No abstract available

Abstract Herein, we introduce a novel methodology to generate urban morphometric parameters that takes advantage of deep neural networks and inverse modeling. We take the example of Chicago, USA, where the Urban Canopy Parameters (UCPs) available from the National Urban Database and Access Portal Tool (NUDAPT) are used as input to the Weather Research and Forecasting (WRF) model. Next, the WRF simulations are carried out with Local Climate Zones (LCZs) as part of the World Urban Data Analysis and Portal Tools (WUDAPT) approach. Lastly, a third novel simulation, Digital Synthetic City (DSC), was undertaken where urban morphometry was generated using deep neural networks and inverse modeling, following which UCPs are re-calculated for the LCZs. The three experiments (NUDAPT, WUDAPT, and DSC) were compared against Mesowest observation stations. The results suggest that the introduction of LCZs improves the overall model simulation of urban air temperature. The DSC simulations yielded equal to or better results than the WUDAPT simulation. Furthermore, the change in the UCPs led to a notable difference in the simulated temperature gradients and wind speed within the urban region and the local convergence/divergence zones. These results provide the first successful implementation of the digital urban visualization dataset within an NWP system. This development now can lead the way for a more scalable and widespread ability to perform more accurate urban meteorological modeling and forecasting, especially in developing cities. Additionally, city planners will be able to generate synthetic cities and study their actual impact on the environment.

Mental health disorders, such as depression, have been estimated to account for the largest proportion of global disease burden. Existing research has established significant correlations between the built environment and mental health. This research, however, has relied on traditional statistical methods that are not amenable to working with large remote sensing image-based datasets. This research addresses this challenge and contributes new knowledge and a novel method for using generative deep learning for urban analysis and synthesis tasks involving mental health. The research specifically investigates three mental state measures: depression, anxiety, and the perception of safety. The experimental results demonstrate the efficacy of the process—providing a new method to find correlational signals, while providing insights on the correlation between specific urban design features and the incidence of depression.

Over the past decade, the field of image-based 3D scene reconstruction and generation has experienced a significant transformation, driven by the integration of deep learning technologies. This shift underscores a maturing discipline characterized by rapid advancements and the introduction of numerous innovative methodologies aimed at broadening research boundaries. The specific focus of this study is on image-based 3D reconstruction techniques applicable to large-scale urban environments. This focus is motivated by the growing need for advanced urban planning and infrastructure development for smart city applications and digitalization, which requires precise and scalable modeling solutions. We employ a comprehensive classification framework that distinguishes between traditional and deep learning approaches for reconstructing urban facades, districts, and entire cityscapes. Our review methodically compares these techniques, evaluates their methodologies, highlights their distinct characteristics and performance, and identifies their limitations. Additionally, we discuss commonly utilized 3D datasets for large environments and the prevailing performance metrics in this domain. The paper concludes by outlining the current challenges faced by the field and proposes directions for future research in this swiftly evolving area.

No abstract available

No abstract available

The problems of inefficient design and difficult data learning faced by conventional building structure design are highly consistent with the advantages of deep learning algorithm learning and inference generation. Intelligent design based on deep learning will be the key technology to solve the design of conventional building structures. This paper takes residential building floor plans as samples, uses a graph neural network model based on graph representation learning, and encodes the samples into graph data structures. By extracting information such as combinations, connections, etc. that are important for a specific objective function, the underlying topological features that form the fundamental building blocks of residential building design are discovered. Combine the discovered building blocks into new designs that meet new user needs and evaluate the feasibility of the resulting candidate solutions.

Conventional methods of planning the layout of a building often involve labor-intensive manual operations that result in solutions that are less than ideal. There is a possibility that the automation of building plan generation could be a game-changer in this business, and the development of deep learning algorithms has opened the door to this possibility. The incorporation of ResN et architecture into the study will assist in overcoming these limitations, which will result in a greater understanding of spatial linkages as well as the capacity to design architectural layouts that are more realistic and varied. The proposed method involves training a deep ResN et model with a dataset consisting of several building layouts. This model is able to capture the intricate interactions that occur between the various architectural components. It is possible to ensure adaptation and flexibility to a variety of design requirements by using the trained model to generate new building plans in accordance with user-defined criteria. In addition to demonstrating that the proposed method is effective, the results also demonstrate how the deep ResN et architecture can deliver a wide range of realistic building layouts.

No abstract available

In the context of the increasing demand for quality living and the rapid modernization of urban areas, residential district layouts face many challenges, such as optimizing functionality within constrained spaces and accommodating the diverse requirements of residents. This study investigates intelligent residential district layout planning methods using deep reinforcement learning to address these difficulties. An agent can learn and make decisions by creating models and structuring residential area components and limits. The deep reinforcement learning method can improve residential layout design schemes by optimizing space utilization, resident convenience, and ecological sustainability. Effective and autonomous data-driven learning is its advantage. Investigating the deep reinforcement learning model and technical techniques for building layout design could improve layout design efficiency and proposal quality. Enhancing urban living quality requires rational, human-centric, and sustainable residential development.

No abstract available

This paper formulates the building layout generation problem and its constraints. The most common solutions are given, the difficulties faced by designers are described. Due to the high labor intensity of classical approaches to solving this problem, the necessity of reviewing models, methods and systems for generating building layout plans is justified. Based on published literature reviews, three main approaches to solving the problem are highlighted: form grammars, genetic algorithms, and deep learning. For each approach the basic concepts and specifics of their use in the problem under consideration are described, comparative tables of features of models and methods in different works are given: types of grammars (context-free, context-sensitive), initial shapes and transformation rules for shape grammars, algorithms (NSGA-II, SPEA-II), populations and optimization criteria for genetic algorithms, architectures (CGAN, CGLO, Pix2Pix, Pix2PixHD) and characteristics of training samples for deep learning (image sizes, conditional channels, rasterization). Dataset preparation in various works devoted to solving the problem of generating building layout plans are described. We systematize the satisfied constraints of the problem in the considered approaches: spatial (site boundaries, street-road network, building heights and layout density) in all cases and architectural (building morphologies, land use types, architectural style and plans of adjacent territories) for genetic algorithms and deep learning models, and compare the reproducibility of solutions on other sites (cross-validation) and visualization capabilities in different works. The key trends and directions for further research highlighted by the authors of the reviewed studies are identified.

The ability of deep learning has been tested to learn graphical features for building-plan generation. However, whether the deeper space allocation strategies can be obtained and thus reduce energy consumption has still not been investigated. In the present study, we aimed to train a neural network by employing a characterized sample set to generate a residential building floor plan (RBFP) for achieving energy reduction effects. The network is based on Pix2Pix, including two sub-models: functional segmentation layout (FSL) generation and building floor plan (BFP) generation. To better characterize the energy efficiency, 98 screened floor plans of Solar Decathlon (SD) entries were labeled as the sample set. The data augmentation method was adopted to improve the performance of the FSL sub-model after the preliminary testing. Three existing residential buildings were used as cases to observe whether the network-generated RBFP gained the effect of decreasing energy consumption with decent space allocation. The results showed that, under the same simulation settings and building exterior profile (BEP) conditions, the function arrangement of the generated scheme was more reasonable compared to the original scheme in each case. The annual total energy consumption was reduced by 13.38%, 12.74%, and 7.47%, respectively. In conclusion, trained by the sample set that characterizes energy efficiency, the RBFP generation network has a positive effect in both optimizing the space allocation and reducing energy consumption. The implemented data augmentation method can significantly improve the network’s training results with a small sample size.

No abstract available

Designing complex, multi-occupancy building layouts requires considering complicated spatial arrangements, design constraints, and strict accessibility requirements. Despite advancements in automatic building layout generation, existing methods struggle to address the complexities of these layouts and often overlook critical accessibility features. This paper presents a novel deep learning-based approach for generating complex, multi-occupancy building layouts that meet both architectural and accessibility standards. To improve the efficiency of training, non-corridor rooms are approximated by minimum rotation rectangles, while a graph neural network (GNN) predicts the number of corners for corridors. To address the quadratic complexity of transformers, we incorporate FlashAttention, to enhance computational efficiency. Accessibility features are integrated into the model by enforcing geometric requirements, including room size ratios and maximum corner distances to account for travel distance to egress. Additionally, a distance penalty is introduced in the loss function to ensure compliance with wheelchair clearance requirements. Experimental results show that our approach outperforms baseline models in generating realistic, complex layouts while ensuring compliance with design and accessibility constraints, making it a robust solution for generating multi-occupancy building layouts.

No abstract available

In recent years, generative design methods are widely used to guide urban or architectural design. Some performance-based generative design methods also combine simulation and optimization algorithms to obtain optimal solutions. In this paper, a performance-based automatic generative design method was proposed to incorporate deep reinforcement learning (DRL) and computer vision for urban planning through a case study to generate an urban block based on its direct sunlight hours, solar heat gains as well as the aesthetics of the layout. The method was tested on the redesign of an old industrial district located in Shenyang, Liaoning Province, China. A DRL agent - deep deterministic policy gradient (DDPG) agent - was trained to guide the generation of the schemes. The agent arranges one building in the site at one time in a training episode according to the observation. Rhino/Grasshopper and a computer vision algorithm, Hough Transform, were used to evaluate the performance and aesthetics, respectively. After about 150 h of training, the proposed method generated 2179 satisfactory design solutions. Episode 1936 which had the highest reward has been chosen as the final solution after manual adjustment. The test results have proven that the method is a potentially effective way for assisting urban design.

In this era, the increasing complexity of sustainable urban development requires intelligent, adaptable architectural design systems that optimize across energy, comfort, and spatial efficiency. This paper describes an Artificial Intelligence (AI) enabled architectural design framework, utilizing the Soft Actor-Critic (SAC) algorithm and a deep reinforcement learning model which is effective for continuous control problems. By modelling layout generation as a Markov Decision Process (MDP), this system allows an autonomous agent to iteratively refine the configuration of a building based on feedback received from environmental simulations. The SAC agent learns to adjust the room layout in the building, room orientation, room material properties, and passive design strategies to maximize a multi-objective reward function based on smart environment objectives. The proposed SAC agent model achieved 30% energy savings, ⩽1.0 Kh/zone thermal discomfort, and 46% operating cost reductions which is significantly better than the Q-learning agent that resulted in only 26.3% savings and a higher level of discomfort.

With the rapid urbanization in China, building energy consumption has become a critical challenge for sustainable urban development. Conventional simulation methods are computationally intensive and inefficient for large-scale urban layout optimization, highlighting the need for fast and reliable predictive approaches. Existing machine learning models often overlook spatial relationships among buildings and rely heavily on manual feature engineering, which limits their applicability at the urban block scale. To address these limitations, the study proposes a building energy consumption prediction model for urban blocks based on Generative Adversarial Networks (GANs), which preserves spatial information while significantly advancing computational speed. The optimal GAN model is further integrated with a Multi-Island Genetic Algorithm (MIGA) to form a GAN-MIGA optimization framework, which is applied to the layout optimization of a target urban block in Lanzhou. Key findings include: (1) the GAN model achieves an average prediction error of 6.8% compared with conventional energy simulations; (2) the GAN-MIGA framework reduces energy consumption by 48.78% relative to the worst-performing solution and by 22.53% compared with the original block layout; (3) the spatial distribution patterns of energy consumption predicted by the GAN are consistent with those obtained from traditional simulation methods; (4) the regression model derived from GAN-MIGA optimization results achieves an R2 value exceeding 0.84; and (5) building layout design strategies are formulated based on key morphological indicators in the regression model. Overall, this study demonstrates the effectiveness of the GAN-based method for urban scale building energy prediction and layout optimization. The proposed GAN-MIGA framework provides practical tools and theoretical support for energy-efficient design, policy formulation, and smart city development, contributing to more sustainable urban energy planning.

No abstract available

No abstract available

In the era of digital intelligence, urban block design methods are transitioning from traditional to automated approaches. As society's demand for spatial quality increases, urban block design methods need to meet more complex requirements. Existing methods have room for improvement in terms of result diversity and dynamic parameter response. To address these challenges, our study develops a rule-based urban block generation design method, formulating the generation rules according to urban characteristics and regulations. Our study utilizes cellular automata and multi-agent algorithms to establish an automated process for urban block generation. Additionally, we develop a multi-objective optimization model to enhance the spatial perception of urban block forms. Our findings reveal that this method effectively addresses the constraints of complexity and dynamically adapts to various parameters. It also successfully generates a diverse array of urban block forms. Overall, our research contributes a novel automated design workflow to the field of computational urban block design, offering novel perspectives and advancing the capabilities of urban planning tools.

PurposeIn many countries, urban block layouts follow rigid and outdated design guidelines; however, these approaches fail to effectively address settlement construction, neighborhood planning, the design of old urban contexts and the rapid development of emerging cities, as seen in the new towns of Iran and various other regions worldwide. Revising these guidelines requires early-stage decision-making to evaluate building layouts based on their impact on the built environment. This study introduces a workflow for optimizing urban block layouts with constant density while considering design constraints, climate and environmental performance. This framework assists planners and policymakers in refining site density distribution and building massing.Design/methodology/approachA computational urban design method was developed utilizing a grasshopper-based algorithm and Python for data generation. It introduces an innovative approach by incorporating buildings’ placement coordinates as design variables alongside dimensions and orientation. The Wallacei tool optimizes layouts by considering energy use intensity (EUI), outdoor thermal comfort and facades’ received radiation employing Ladybug Tools. Spatial daylight autonomy (sDA) and urban heat island (UHI) effects are also integrated into the workflow. The framework is applied to a hypothetical residential district in Tehran, consisting of nine blocks, each containing nine plots.FindingsThe results for a set of optimal solutions demonstrate that beyond existing urban design policies, environmentally prioritized designs are achievable. Radiation differences between the optimal models and the conventional parallel layout range from 41.3 to 61.1%. However, changes in EUI and universal thermal climate index (UTCI) are minimal; significant improvements are observed in sDA, with 7, 8 or all 9 buildings meeting the acceptable threshold compared to none in the parallel layout. These findings highlight the effectiveness of the design method framework proposed in this research, which has the potential to significantly influence urban planning practices across various regions globally.Originality/valueThe methodology of this study provides a flexible decision-making framework for early design phases, enabling policymakers, architects and urban designers to address the gap between current urban design methods and environmentally driven approaches. Sensitivity analysis of optimal solutions further offers valuable insights into the correlation between design variables and evaluation metrics.

The iterative process of urban development often produces fragmented renewal zones that disrupt the continuity of urban morphology, undermining both cultural identity and economic cohesion. Addressing this challenge, this study proposes a generative design framework based on Deep Convolutional Generative Adversarial Networks (DCGANs) to predict and regenerate urban morphology in alignment with existing spatial contexts. A dataset was constructed from highly urbanized city centers and used to train a DCGAN model. To evaluate the model performance, seven landscape pattern indices—LPI, PLAND, LSI, MPFD, AI, PLADJ, and NP—were employed to quantify changes in scale, shape, compactness, fragmentation, and spatial adjacency. Results show that the model accurately predicts morphological patterns and captures underlying spatial logic in developed urban areas, demonstrating strong sensitivity to local form characteristics, and enhancing the feasibility of sustainable urban renewal. Nonetheless, the model’s generalizability is constrained by inter-city morphological heterogeneity, highlighting the need for region-specific adaptation. This work contributes a data-driven approach to urban morphology research and offers a scalable framework for form-based, sustainability-oriented urban design.

No abstract available

In modern urbanism, (re)production of urban land predominantly relies on large parcels through intensive capital investments. Such a mainstream significantly shapes the overall urban form, subsequently influencing the quality of life through the perceived characteristics of the form and program of the planned districts. Consequently, critical urban design theory increasingly prioritizes the plot as the fundamental unit of future urban development. While ‘plot-based urbanism’ presents a responsive approach to this issue, there remains a notable gap in systematic methodologies that can be universally applied across different contexts. In this paper, the authors propose an algorithmic framework that would be employed as a design control tool based on the associative logic of plot-based urban formation. The model framework comprises three steps: (1) plot layout generation, (2) building configuration, and (3) incremental formation of the block fabric. The applied model demonstrates the compositional variation and coherence within the urban block while concurrently optimizing the climatic performance of the emerging fabric.

In recent years, various modeling efforts have been undertaken to design and plan buildings for increasing lo-cal renewable energy production. To this end, advanced generative models have been developed to explore various urban forms that may provide better access to renew-ables. However, current approaches have failed to adequately evaluate the effectiveness of a design from a lo-cal energy network performance and a collective energy efficiency perspective. An auto-decoder model has been developed to generate planning envelopes, and configurations are evaluated based on integrated local energy networks. The optimal form is identified, which maximizes collective energy efficiency at the neighborhood level.

The urban design process is complex and interdisciplinary, especially in the context of high-density cities with high-rise buildings. The design of high-rise buildings requires input from a variety of stakeholders in the city, who often represent conflicting requirements and interests. However, conventional design approaches struggle to address this complexity. This study introduced a generative urban design approach and applied it to a case study for high-rise urban development in Guangzhou, China. Specifically, 7500 urban forms were generated with variables such as street networks, block offset, building typology, and green space, and then were evaluated and optimized via multi-objective evolutionary algorithms. A total of 30,000 performance values were generated through various simulations. This study also presented a two-round, quantitative evaluation process against eight performance objectives from environmental, social, and economic aspects, including urban density, green space area, Daylight Potential, construction cost, Heat Stress, Green Space Accessibility, View Interest, and Shadow Impact.

The study of performance-driven optimization (PDO) in urban block design is essential in the context of architectural form and urban sustainability. PDO focuses on the integrated and comprehensive optimization of various quantifiable performances of buildings, such as solar energy usage, thermal comfort, and energy efficiency. This method aligns urban spaces with sustainable development principles, ensuring they are not only aesthetically pleasing but also functionally efficient. This study explores the existing deficiency in the literature by conducting an in-depth scientometric analysis of PDO in urban block design. Employing science mapping coupled with bibliometric analysis using Python, this study meticulously analyzes the prevailing literature to map out the current intellectual landscape, understand trends, and identify key themes within this domain. This review identifies the key trends, methodologies, and influential works shaping the dynamic field of PDO. It emphasizes the critical roles of computational simulation, artificial intelligence integration, and big data analytics in refining urban block design strategies. This study highlights the growing importance of energy efficiency, environmental sustainability, and human-centric design elements. This review points to an increasing trend towards using sophisticated modeling techniques and data-driven analysis as essential tools in urban planning, crucial for developing sustainable, resilient, and adaptable urban spaces.

No abstract available

Abstract The significant challenges that urban design faces when moving forward into this new digital era involve the increasingly massive and complex projects that must be analyzed and developed. The need for efficient analysis tools and rational design methods represent ongoing concerns, while practicable and maneuverable applications rather than experimental studies are scarce. This research implemented a framework of digital description and generative grammar of block form from the perspective of block morphological complexity. The implemented framework was tested in an urban design practice. First, this study adapted the hierarchical structure and access structure theory to analyze the spatial form of target blocks. The characteristics of blocks case samples in Nanjing are extracted as text descriptions. Second, the relevant composition patterns and parameters were employed as features for classification and were converted into procedural rules. With the top-down control by rules and bottom-up generation by shape grammar in CityEngine, the texture of the block can be generated as close to the actual block as possible. Furthermore, in a real urban design case located in Nanjing, this work applies the method to construct a three-dimensional scene quickly and accurately. After integrating design factors such as an environment, transportation, and vision and summarizing the intentions of blocks and buildings in corresponding functions and control indexes, the initial generation plan was built by applying the obtained characteristics and procedural rules in specific shape grammar. Finally, designers can adjust the result in detail by employing real-time calculation and interactive operation.

Public spaces play a crucial role in urban environments, and their design must consider various factors to optimize functionality and user experience. CivicMorph, a generative modeling framework, stands at the forefront of this challenge, utilizing computational design and generative algorithms to craft responsive urban spaces. It emphasizes inclusivity by incorporating user-centric data alongside environmental parameters to create diverse spatial configurations. The framework fosters collaboration among stakeholders through design patterns and spatial analytics, ensuring that the planning process is participatory. Case studies illustrate CivicMorph's capability to produce public space designs that adapt to social dynamics and environmental considerations. The findings highlight notable enhancements in user satisfaction and functional efficiency, positioning CivicMorph as an innovative solution for urban design challenges.

Optimizing urban morphology is pivotal for lowering energy consumption and carbon emissions while supporting sustainable development. However, traditional methods often rely on inefficient post-evaluation, hindering direct, goal-oriented design feedback. This study thus presents the PC-GAN model, an automated block design tool that integrates Pix2pix with a two-step GAN mechanism inspired by CycleGAN, generating high-resolution 3D urban forms under specific climatic and energy-reduction targets. Focusing on Guangzhou’s six main LCZ types (LCZ1, LCZ2, LCZ4, LCZ5, LCZ6, LCZ8) and iterating 10% stepwise reductions in baseline energy use, the model illuminates distinctive low-carbon optimization pathways across various urban morphologies. In doing so, PC-GAN offers rapid generation of diverse design alternatives while fulfilling low-energy requirements. This new framework not only improves planning efficiency and flexibility but also directs future research toward more adaptive, data-driven methodologies for low-carbon urban morphology optimization.

ABSTRACT Integrating deep generative models into urban form generation is an innovative and promising approach to support the urban design process. However, most deep generative urban form models are based on image representations that do not explicitly consider topological relationships among urban form elements. Toward developing an urban form generation framework aided by deep generative models and considering topological information, this paper reviews urban form generation, deep generative models/deep graph generation, and the state of the art of deep generative models in architectural and urban form generation. Based on the literature review, a topology-based urban form generation framework aided by deep generative models is proposed. The hypotheses of street network generation by deep generative models for graph generation and plot/building configuration generation by deep generative models/space syntax and the feasibility of the proposed framework require validation in future research.

Community roads are crucial to community navigation. There are automatic methods to obtain community roads using trajectories, but the sparsity and uneven density distribution of community trajectories present significant challenges in identifying community roads. To overcome these challenges, we propose a conditional generative adversarial network (MAC-GAN) supervised by pedestrian trajectories and neighborhood building footprints for road generation. MAC-GAN packs the “road trajectory–building footprint” pairs into images to characterize implicit ternary relations and sets up a multi-scale skip-connected and asymmetric convolution-based generator to incorporate such a relationship, in which the generator and discriminator mutually learn to optimize the network parameters and then derive approximate optimal results. Experiments on 37 real-world community datasets in Wuhan, China, are conducted to verify the effectiveness of the proposed model. The experimental results show that the F1 score of our model increases by 1.7–6.8%, and the IOU of our model increases by 2.2–7.5% compared with three baselines (i.e., Pix2pix, GANmapper, and DLinkGAN (configured by DLinknet)). In areas with sparse and missing trajectory data, the generated fine roads have high accuracy with the supervision of building footprints.

Generative Adversarial Network (GAN) is widely used in many generative problems, including in spatial information sciences and urban systems. The data generated by GANs can achieve high quality to augment downstream training or to complete missing entries in a dataset. GANs can also be used to learn the relationship between two datasets and translate one into another, e.g. road network data into building footprint data. However, such approach has not been developed in the geospatial and urban data science context, its usability remains unknown, and the methods are not fully developed. We develop a new Geographical Data Translation algorithm based on GAN to generate high-resolution vector building data solely from street networks, which may be used to predict the urban morphology in absence of building data, also enabling studies in unmapped or undermapped urban geographies, among other advantages. Experiments on 16 cities around the world demonstrate that the generated datasets are largely successful in resembling ground truth morphologies. Thus, the approach may be used in lieu of traditional data for tasks that are often hampered by lack of data, e.g. urban form studies, simulation of urban morphologies in new contexts, and spatial data quality assessment. Our work proposes a novel rapid approach to generate building footprints in replacement of procedural methods and it introduces a new intrinsic method for large-scale spatial data quality control, which we test on OpenStreetMap by predicting missing buildings and suggesting the completeness of data without the usually required authoritative counterparts. The code, sample model, and dataset are available openly.

In architecture, landscape architecture, and urban design, site planning refers to the organizational process of site layout. A fundamental step for site planning is the design of building layout across the site. This process is hard to automate due to its multi-modal nature: it takes multiple constraints such as street block shape, orientation, program, density, and plantation. The paper proposes a prototypical and extensive framework to generate building footprints as masterplan references for architects, landscape architects, and urban designers by learning from the existing built environment with Artificial Neural Networks. Pix2PixHD Conditional Generative Adversarial Neural Network is used to learn the mapping from a site boundary geometry represented with a pixelized image to that of an image containing building footprint color-coded to various programs. A dataset containing necessary information is collected from open source GIS (Geographic Information System) portals from the city of Boston, wrangled with geospatial analysis libraries in python, trained with the TensorFlow framework. The result is visualized in Rhinoceros and Grasshopper, for generating site plans interactively.

Building footprint generation is a vital task of satellite imagery interpretation. However, the segmentation masks of buildings obtained by existing semantic segmentation networks often have blurred boundaries and irregular shapes. In this research, we propose a new boundary regularization network for building footprint generation in satellite images. More specifically, we consider semantic segmentation and boundary regularization in an end-to-end generative adversarial network (GAN). The learned building footprints are regularized by the interplay between the generator and discriminator. By doing so, the straight boundaries and geometric details of the building could be preserved. Experiments are conducted on a collected dataset of Planetscope satellite imagery (spatial resolution: 4.77 m/pixel). Our approach is much superior to the state-of-the-art methods in both quantitative and qualitative results.

This study is committed to using generative adversarial network to achieve collaborative optimization of building surface morphology design and energy consumption computer simulation, and proposes an improved algorithm of generative adversarial network based on multimodal information fusion. The research team collected case data of 50 building projects from different regions, functions and scales, covering meteorological data, building material parameters, and multimodal information such as building surrounding environment images and functional requirements documents required for energy consumption simulation. The experimental setting is a comparison group, using traditional generative adversarial network and improved algorithm to generate building surface morphology, and energy consumption simulation is performed on the generated results. The results show that the information entropy index of the building surface morphology generated by the improved algorithm is improved by 30% in terms of diversity, and the degree of fit with design requirements is increased from 60% to 85% after quantitative evaluation. The energy consumption simulation results show that the annual energy consumption of the building generated by the improved algorithm is reduced by an average of 1500kWh and the peak energy consumption is reduced by 10kW, which is significantly superior to the traditional algorithm in energy consumption control. In terms of algorithm performance, the convergence speed of the improved algorithm increased by 40% and the computing resource consumption decreased by 25%.

No abstract available

Generative AI technologies have rapidly advanced in recent years, finding extensive applications in architectural design, including the generation of building volumes, which represent the three-dimensional form of buildings. However, current studies in this area lack scientific dataset classification methods and primarily focus on single generative models, with limited emphasis on model integration. To address these challenges, this study presents a comprehensive workflow that integrates multiple spatial computation methods to calculate the feature data of building volumes in urban areas, followed by cluster analysis to classify the building volume dataset. A generative model, combining Conditional Generative Adversarial Networks (CGAN) and Denoising Diffusion Probabilistic Models (DDPM), is introduced to improve the accuracy and diversity of generated building volume point clouds. The workflow is validated through a case study of the old town of Kashgar, demonstrating its effectiveness and robustness. Results show that the trained model achieves a mean training accuracy of 93.88% and a mean validation accuracy of 89.26%, with the training loss stabilizing at approximately 0.1 after 176 epochs. This approach accurately classifies the building volume dataset and generates diverse 3D point clouds of building volumes, highlighting its potential for the generative design of building masses in complex urban neighborhoods. This study offers new strategies for enhancing the design process using AI technologies.

In recent years, generative Artificial Intelligence (AI) technologies have gradually matured in the field of collaborative architectural design, providing innovative technical pathways for the protection and inheritance of architectural cultural heritage. The Lingnan arcade building (Qilou), as a characteristic architectural form in Southern China, faces significant challenges regarding its digital preservation and urban renewal. This paper proposes a quantitative and technical framework for applying the Stable Diffusion (SD) model to the architectural scheme generation of Lingnan arcade buildings. By establishing a systematic workflow comprising text-to-image, image-to-image, and line-drawing-to-image generation modes, we conduct small-sample empirical experiments based on real-world arcade building data. The generated schemes are quantitatively evaluated using Structural Similarity Index Measure (SSIM), Frechet Inception Distance (FID), and Contrastive Language-Image Pre-training (CLIP) scores. Experimental results demonstrate that the image-to-image generation mode achieves the highest structural fidelity (SSIM $=0.82)$ and the best overall visual quality $(\mathbf{F I D}=\mathbf{4 5. 3})$, effectively preserving historical features while integrating modern design elements. The proposed framework provides an efficient, verifiable, and reproducible technical paradigm for the digital preservation and adaptive reuse of regional architectural heritage, significantly reducing the manual modeling workload in urban renewal projects.

No abstract available

Preserving the architectural heritage of traditional historical districts is a crucial aspect of urban renewal. Traditional design workflows are time-consuming and subjective. Current data-driven design methods generate specific style images but are labor-intensive. Recent research highlights Stable Diffusion models’ potential in generating high-fidelity images based on prompts. However, research applying these models to historical districts is scarce, with challenges in creating effective prompts and training parameters. This study proposes a framework combining Stable Diffusion models with expert system-based techniques to generate architectural facades from professional prompts. We constructed a dataset of traditional arcade facades, trained Low-Rank Adaptation (LoRA) models, and integrated ControlNet models to enhance controllability. Experimental results showed our models excelled in precision, realism, and diversity. Both qualitative and quantitative evaluations, along with practical application tests, confirmed our approach aids designers and prompts engineers, contributing to the preservation of architectural heritage and the renewal of urban historical districts.

No abstract available

The urban road spatial structure is a crucial and complex component of urban design. Generative design models, such as the Stable Diffusion model, can rapidly and massively produce designs. However, the opacity of their internal architecture and the uncertainty of their outcomes mean that the results generated do not meet specific disciplinary assessment criteria, thus limiting their widespread application in planar design and planning. Additionally, traditional software processes targeting specific indicators are time-consuming and do not allow for rapid evaluation. To address these challenges, we utilized several areas of the road spatial structures in six cities and their corresponding four space-syntax parameters as training samples. We simultaneously trained two models: one is a LoRA Model based on the Stable Diffusion architecture used for generating road networks similar to those of various city road spatial structures; the other is a CoAtNet Model (Convolution + Transformer) used as an evaluation model to predict the space-syntax parameters of road structures and calculate the Mean Absolute Percentage Error (MAPE) relative to real urban samples. Subsequently, by linking these two models end-to-end, we were able to filter out generated samples with the smallest MAPE, thereby enhancing the structural similarity between the generated results and the actual urban road spatial structures. This process of rapid generation and swift evaluation of network configurations marks a critical advancement towards better performance and more customized design solutions.

No abstract available

Predicting urban morphology based on local attributes is an important issue in urban science research. The deep generative models represented by generative adversarial network (GAN) models have achieved impressive results in this area. However, in such methods, the urban morphology is assumed to follow a specific probability distribution and be able to directly approximate the distribution via GAN models, which is not a realistic strategy. As demonstrated by the score-based model, a better strategy is to learn the gradient of the probability distribution and implicitly approximate the distribution. Therefore, in this paper, an urban morphology prediction method based on the conditional diffusion model is proposed. Implementing this approach results in the decomposition of the attribute-based urban morphology prediction task into two subproblems: estimating the gradient of the conditional distribution, and gradient-based sampling. During the training stage, the gradient of the conditional distribution is approximated by using a conditional diffusion model to predict the noise added to the original urban morphology. In the generation stage, the corresponding conditional distribution is parameterized based on the noise predicted by the conditional diffusion model, and the final prediction result is generated through iterative sampling. The experimental results showed that compared with GAN-based methods, our method demonstrated improvements of 5.5%, 5.9%, and 13.2% in the metrics of low-level pixel features, shallow structural features, and deep structural features, respectively.

Generative modeling of urban street networks poses a distinct challenge in geometric deep learning. It requires synthesizing large-scale graphs that respect physical constraints while capturing statistical patterns shared across real-world cities. Existing approaches often struggle to balance global spatial coherence with precise topological connectivity. Furthermore, many are limited to modifying reference templates rather than generating diverse, novel structures from scratch. In this work, we propose a scalable two-stage framework that effectively decouples these modalities. We employ an autoregressive Transformer to learn diverse node distributions from coordinate sequences, and introduce SCOPE, a specialized edge predictor that conditions on global spatial context while respecting local geometric constraints. Trained on over 35,000 cities, our approach outperforms baselines with 0.91 average precision in edge prediction and faithfully reproduces high-order graph statistics. Qualitatively, it synthesizes physically plausible networks, demonstrating the capacity to hypothesize novel urban forms beyond the training distribution. Additionally, we show that the learned geometric embeddings intrinsically capture spatial laws, offering interpretability alongside generative performance.

This research is primarily focused on utilizing available airborne LiDAR data and spatial data from the OpenStreetMap (OSM) database to generate 3D models of buildings for a large-scale urban area. The city center of Ljubljana, Slovenia, was selected for the study area due to data availability and diversity of building shapes, heights, and functions, which presented a challenge for the automated generation of 3D models. To extract building heights, a range of data sources were utilized, including OSM attribute data, as well as georeferenced and classified point clouds and a digital elevation model (DEM) obtained from openly available LiDAR survey data of the Slovenian Environment Agency. A digital surface model (DSM) and digital terrain model (DTM) were derived from the processed LiDAR data. Building outlines and attributes were extracted from OSM and processed using QGIS. Spatial coverage of OSM data for buildings in the study area is excellent, whereas only 18% have attributes describing external appearance of the building and 6% describing roof type. LASTools software (rapidlasso GmbH, Friedrichshafener Straße 1, 82205 Gilching, GERMANY) was used to derive and assign building heights from 3D coordinates of the segmented point clouds. Various software options for procedural modeling were compared and Blender was selected due to the ability to process OSM data, availability of documentation, and low computing requirements. Using procedural modeling, a 3D model with level of detail (LOD) 1 was created fully automated. After analyzing roof types, a 3D model with LOD2 was created fully automated for 87.64% of buildings. For the remaining buildings, a comparison of procedural roof modeling and manual roof editing was performed. Finally, a visual comparison between the resulting 3D model and Google Earth’s model was performed. The main objective of this study is to demonstrate the efficient modeling process using open data and free software and resulting in an enhanced accuracy of the 3D building models compared to previous LOD2 iterations.

With the rapid development of computer graphics technology, procedural modeling is a method to efficiently generate virtual 3D models, and complex geometric models are generated through certain algorithm rules. Traditional procedural modeling techniques have a high degree of automation, but the rule design is cumbersome, and it is difficult to adapt to different styles or highly customized architectural needs, and the flexibility and editability are limited. In this paper, a procedural generation method of urban building single model based on template matching algorithm is proposed, which introduces CGA (Computer Generated Architecture) shape grammar to guide the layout generation of building facade structure. The creation process of the model is to create a freeform white mold as the building basic model, then create the building facade with the custom constraint layout of the subsidiary structure, and finally combine the white mold and the facade structure through the template matching algorithm proposed in this paper. The entire creation process is parametrically controlled, mapping from the building layer to the building parameters, and by adjusting the parameters, it is possible to generate a 3D building model that meets the needs and has a variety of appearances and is realistic. Experimental results show that this method can realize rapid building modeling and significantly improve the modeling efficiency.

This paper presents an approach to generate interiors of known real buildings using target parameters. This can be useful in simulations in hazard situations, earthquake simulations or architecture, archaeology, engineering and defense contexts. However, modeling this type of complex structure requires a significant amount of time and manual work, especially when it comes to designing and planning the interior according to important parameters, such as room size and ratio. This paper presents a method for procedurally generating interior floor plans, based on buildings extracted from Geographic Information Systems (GIS). This approach subdivides a building into rooms and corridors, parametrized by room area and room ratio, based only on its geometry. This method is designed to adapt to imperfect data with an unspecified orientation. To evaluate this approach, we conducted a user study with 29 participants. The results show that this approach is particularly strong in how it handles imperfect information and manages space and is stronger at generating smaller rooms. This method provides a fast and effective way to generate building interiors, independently of their orientation or the existence of imperfect data, with a focus on the study of shock wave propagation through them.

In building massing design, using passive design strategies is a critical approach to reducing energy consumption while offering comfortable indoor environments. However, it is often impractical for architects to systematically explore passive design strategies at the outset of the building massing design and architectural form-finding processes, which may result in inefficient or ineffective utilization of the strategies. To address this issue, this study presents a reverse passive design strategy exploration approach that leverages the capability of computational optimization and parametric modeling to help architects identify feasible passive design strategies for building massing design. The approach is achieved using a building massing design generation and optimization tool, called EvoMass, and various building performance simulation tools in Rhino-Grasshopper. The optimization can produce site-specific design references that reflect rich performance implications associated with passive design strategies, such as atriums and self-shading. As such, architects can screen out promising passive design strategies corresponding to different performance factors from the optimization result. Two case studies related to daylighting, sky exposure, and solar heat utility are presented to demonstrate the approach, and the relevant utility and limitations are discussed.

No abstract available

Procedural modeling offers significant advantages over traditional methods of geovisualizing 3D building models, particularly in its use of scripts or machine language for model description. This approach is highly suitable for computer processing and allows for the rapid rendering of entire building models and cities, especially when the buildings are not highly diverse, thus fully leveraging the strengths of procedural modeling. The first hypothesis is that buildings in the real world are mostly different and they should still be able to be displayed through procedural modeling procedures, and the second hypothesis is that this can be achieved in several ways. The first hypothesis suggests that real-world buildings, despite their diversity, can still be effectively represented through procedural modeling. The second hypothesis explores various methods to achieve this representation. The first approach involves recognizing the basic characteristics of a building from photographs and creating a model using machine learning. The second approach utilizes artificial intelligence (AI) to generate detailed building models based on comprehensive input data. A script is generated for each building, making reverse procedural modeling in combination with AI an intriguing field of study, which is explored in this research. To validate this method, we compare AI-generated building models with manually derived models created through traditional procedural modeling techniques. The research demonstrates that integrating AI and machine learning techniques with procedural modeling significantly improves the efficiency and accuracy of generating 3D building models. Specifically, the use of convolutional neural networks (CNNs) for image-to-geometry translation, and Generative Adversarial Networks (GANs) for texture generation, showed promising results in creating detailed and realistic 3D structures. This research is significant as it introduces a novel methodology that bridges the gap between traditional procedural modeling and modern AI-driven techniques. It offers a robust solution for automated 3D building modeling, potentially revolutionizing the fields of urban planning and architectural design by enabling more efficient and accurate digital representations of complex building geometries.

The early-stage performance of architectural design is critical for long-term carbon optimization. This study investigates three constructive solid geometry (CSG) strategies, namely union, subtraction, and intersection, within a parametric workflow driven by the Non-dominated Sorting Genetic Algorithm II (NSGA-II), aiming to minimize operational carbon emissions (OCE) and maximize gross floor area (GFA). It connects geometric design logic with objectives to support carbon-conscious massing decisions. The study examines how CSG strategies influence simulation runtime, model validity, and design diversity; how NSGA-II optimizes form generation under regulatory constraints e.g., floor area ratio (FAR), open space ratio (OSR); and how these affect early-stage outcomes. Out of 3,000 generated configurations, the subtraction strategy achieved the lowest operational carbon intensity (181.30 kgCO₂eq/m² at 45,901.85 m² GFA) and provided greater design diversity with 151 valid configurations, while the intersection strategy demonstrated the fastest simulation runtime of 3h 40m 40s. The proposed framework effectively filtered infeasible designs and ensured regulatory compliance to maintain model validity. Convergence of optimal OCE values (~181 kgCO₂eq/m²) across strategies suggests a useful benchmark for high-rise office buildings in dense urban contexts. These findings highlight the value of integrating carbon optimization, regulatory constraints, and multi-objective parametric design into early-stage architectural workflows to improve sustainability.

No abstract available

In architecture, context term refers to the surrounding area in which the building is situated. Context is crucial in architectural theory and practice when developing architectural vocabulary. Creating harmony with historic context amidst expanding development has been a primary concern and focus of designers since the mid-twentieth century. Contextualism theory refers to the interaction between new buildings and their surroundings, as well as the challenge of integrating new and existing structures to establish congruence and continuity. When using a contextual design method, an analysis of traditional architectural style has a considerable impact on a designer’s decision-making process (Lambe and Dongre, 2019). In this study, a general framework is proposed, which integrates shape grammar-based procedural modeling and augmented reality technology to create a harmonious environment through the generation of new designs based on the grammar of the existing architectural style of the residential buildings in Amman city without limiting the designer’s creativity of order to address the issue of Western architectural movements influencing the architectural style in Amman, which is incompatible with the city’s identity and it’s context. This approach was tested on a group of University of Jordan students, and the results were evaluated using a machine learning model. The effectiveness of this method was discovered, and it can be regarded as a step toward achieving spatial congruence.

: In this paper, we present an approach to model spatial influences in multi-agent models using procedural generation and genetic algorithms. We applied this approach in an urban growth model. In agent-based simulations, the agents make decisions based on the perception of their environment. In our context, the agents represent inhabitants who can create new buildings or extend the existing ones. Their behaviour is ruled by spatial influences (e.g., the proximity of the road increases chances of building in the surrounding areas). Procedural generation provides a good framework for representing the influences of the environment on the agent’s behaviour. Each spatial feature is associated with an influence function. Their parameters search space can be tremendous, making it difficult for field experts to set them manually. Consequently, we use a genetic algorithm to optimize the parameters of these influence functions and train the model based on three spatial measures (Chamfer distance, kernel density, and a density grid). This approach can be employed likewise to any problem where the agent decisions are wholly or partly based on location. Our experiments highlight the interest of our approach and the impact of the chosen fitness functions.

Within the framework of the Healthy China strategy, daylighting in primary and secondary schools is crucial for students’ health and learning efficiency. Most schools in China still face insufficient and uneven daylighting, along with limited outdoor solar exposure, underscoring the need for systematic optimization. Guided by the “Daylighting School” concept, this study proposes a campus design model that integrates indoor daylighting with outdoor activity opportunities and explores a generative optimization approach. The research reviews daylighting and thermal performance metrics, summarizes European and American “Daylighting School” experiences, and develops three classroom prototypes—Standard Side-Lit, High Side-Lit, and Skylight-Lit—together with corresponding campus layout models. A two-stage optimization experiment was conducted on a high school site in Guangzhou. Stage 1 optimized block location and functional layout using solar radiation illuminance and activity accessibility distance. Stage 2 refined classroom configurations based on four key performance indicators: sDA, sGA, UOD, and APMV-mean. Results show that optimized layouts improved activity path efficiency and daylight availability. High Side-Lit and Skylight-Lit classrooms outperformed traditional Side-Lit in illuminance, uniformity, and glare control. To improve efficiency, an ANN-based prediction model was introduced to replace conventional simulation engines, enabling rapid large-scale assessment of complex classroom clusters and providing architects with real-time decision support for daylight-oriented educational building design.

Achieving robust environmental responsiveness in early stage architectural spatial layout design remains a critical challenge under the imperatives of global carbon neutrality and climate-adaptive building practice. Conventional parametric design and multi-objective optimization approaches suffer from computational inefficiency, inadequate constraint satisfaction, and opaque generative logic when operating in high-dimensional design spaces. This paper presents a mathematically rigorous, climate-responsive spatial layout generation framework that unifies category theory with conditional diffusion modeling. The proposed method formalizes site-specific environmental parameter systems and architectural spatial topologies as two small categories, and establishes structure-preserving environment-to-space mappings via covariant functors; natural transformations are further introduced to characterize morphological transitions across distinct design strategies. A conditional diffusion model (CDM) serves as the generative engine, producing candidate spatial topological configurations subject to environmental parameter conditioning. A three-stage categorical constraint screening mechanism—constructed from groupoid structures and pullback limits—enforces simultaneous compliance with functional adjacency requirements, topological coherence, and multi-criteria environmental performance targets. Extensive experiments across three climatically contrasting sites (Hangzhou, Qingdao, and Lijiang) demonstrate that the framework substantially enhances environmental response performance while preserving spatial topological rationality, achieving competitive generation efficiency and constraint satisfaction relative to conventional parametric optimization baselines. These findings establish that categorical structures can serve as interpretable, mathematically consistent constraint engines within AI-driven generative design pipelines, offering a principled computational paradigm for climate-responsive architectural layout synthesis.

No abstract available

No abstract available

Objective . The purpose of the study is to explore the possibilities of optimizing the layout of construction facilities using modern computer technologies and algorithms, such as generative design. Method . Research methods include generative design and machine learning algorithms. Result. The results of the development of a methodology for optimizing the layout of construction facilities, which can be applied in the practical activities of architects and engineers, are presented. The principles of information modeling help to increase the speed of project work without compromising its quality. This helps to prevent a lot of errors at the design stage, which in turn reduces the likelihood of problems during the construction phase of the facility. Conclusion . The complexity of the layout optimization task increases with the increasing complexity of the project and the variety of requirements that the facilities must meet. Using the Dynamo Studio development environment will allow you to create an effective software module for generating the layout of construction objects in space.

No abstract available

Driven by the two goals of global carbon neutrality and improvement of living standards, indoor environments must respond to both the functional rationality of space and the efficient use of energy. Problem: Existing design solutions often suffer from low space use, high energy consumption of HVAC systems (heating, ventilation and air conditioning), low adaptability to user needs, and chronological optimization processes with low iteration efficiency. This article first builds a technical data collection model training framework for optimizing multi-target generation verification, with an enhanced generative adversary network (GAN) as the basic AI generative algorithm. data on equipment energy consumption and user preference characteristics to establish a multi-purpose optimization model that balances configuration rationality and energy efficiency indicators; Secondly, through simulation experiments and real case tests, model parameters are optimized and effectiveness verified. The experimental study shows that the method proposed in this article has higher scores than traditional manual design in all dimensions, the most important improvements being in functional adaptation (4.6 points vs. 3.2 points) and comfort (4.5 points vs. 3.1 points).

Generative design based on machine learning has become an important area of application for artificial intelligence. Regarding the generative design process for residential site plan layouts (hereafter referred to as “RSPLs”), the lack of experimental demonstration begs the question: what are the design preferences of machine learning? In this case, all design elements of the target object need to be extracted as much as possible to conduct experimental studies to produce scientific experimental results. Based on this, the Pix2pix model was used as the test case for Chinese residential areas in this study. An experimental framework of “extract-translate-machine-learning-evaluate” is proposed, combining different machine and manual computations, as well as quantitative and qualitative evaluation techniques, to jointly determine which design elements and their characteristic representations are machine learning design preferences in the field of RSPL. The results show that machine learning can assist in optimizing the design of two particular RSPL elements to conform to residential site layout plans: plaza paving and landscaped green space. In addition, two other major elements, public facilities and spatial structures, were also found to exhibit more significant design preferences, with the largest percentage increase in the number of changes required after machine learning. Finally, the experimental framework established in this study compensates for the lack of consideration that all design elements of a residential area simultaneously utilize the same methodological framework. This can also assist planners in developing solutions that better meet the expectations of residents and can clarify the potential and advantageous directions for the application of machine learning-assisted RSPL.