Abstract
Applying Machine Learning (ML) has seen rapid progress in many disciplines, such as architectural design. Recent research reveals promising potential for integrating ML in solving design problems. This paper explored how ML can serve as a tool to guide designing action. It conducted thematic analyses of ML experiments in the architecture domain to build a framework addressing two key aspects: the tasks achieved and the required training dataset. The paper found that ML mainly performs design generation, optimization, and recognition via classification and clustering. Three scenarios of design generation using ML have started from pre-design requirements and constraints, conceptual design, or parametric design. ML can predict design features based on prescribed performance or calculate performance metrics, based on varied design options. Design recognition classifies or clusters designs to detect their styles, typologies, and morphologies, besides tracking the process of best practices. The specifications of training datasets vary in terms of data sources as real or virtual, techniques for representing datasets as visual or textual, and the relationships between input and output datasets as refinement, embodiment, sorting, or evaluation. The findings revealed that ML has a wide range of experimentation and opens more opportunities for further integration in the design process.
1 Introduction
Instrumentalization of Machine Learning (ML) investigates how this method can be applied as an instrument to guide the design tasks in the architectural design process. ML is an artificial intelligence method that adopts system learning from past experiences. It builds an algorithmic model to describe data, predict outcomes, and make decisions based on the training dataset. It can generate a large space of design outputs in a short time [1].
ML workflow includes three sequential processes: training, testing, and assessment. The computational method of ML learns from data in the training and testing phases by identifying data mapping between the input factors and the output solution [2]. The quality and quantity of training datasets affect the reliability of the machine-learning outcomes [3, 4]. Consequently, the training process of ML in architectural design requires large datasets to teach, test, and validate the algorithmic model [5]. This process analyzes the design qualities and the relations between the pair of input and output databases such as geometric features, dimensions, orientation, etc.
During the training phase, the input data is interpreted, detected, and processed through a feedback cycle in a learning process to improve the output data [6]. The training process can be controlled by adjusting dataset parameters. Different training datasets for the same design task make the models of ML perform differently.
After training the algorithmic model, the testing process receives new design inputs that were not included in the training process to validate the algorithmic model. This process utilizes the identified qualities and relations of the training dataset to generate the design output. The predicted output is based on the design pattern of the training dataset [5]. The third assessment process aims to verify the learning ability of the ML algorithmic model during the training process by evaluating the output of the testing process. For each ML experiment, specific design criteria and constraints are used to evaluate the predicted output and to assess the learning process of the algorithmic trained model [1].
ML has three types of learning strategies: “supervised, unsupervised, and reinforcement learning”. The most common of them are “supervised and unsupervised learning”. The training data in “supervised learning” is labeled and has predefined solutions. After training, the system is subject to test data to investigate its feasibility for future predictions. In contrast, training data in unsupervised learning is unlabeled and does not have predefined solutions. The algorithms attempt to structure the input data into defined clusters.
This paper explores the capabilities of ML system applications in architectural design. It is structured in the following sections. Section 2 reviews previous studies on ML experiments in architecture. Section 3 describes the research aim and methodology. A framework for ML tasks and datasets is presented in section 4. The last section is the discussion and conclusion.
2 Machine learning involvement in architectural design
ML is a “trial and error” learning process that captures the description of architectural features. Karabagli et al. applied ML to generate three-dimensional volumetric models based on an input dataset of graph representations translated into 3D massing model. They implemented HouseGAN++ to convert graph nodes and edges that encode rooms and their spatial adjacency into floorplans. Then, the generated floorplans were processed to identify rooms, walls, doors, and windows and to generate a 3D volumetric model [7].
Boim et al. implemented ML models in the urban design generation of non-planned settlements based on a training dataset of existing local patterns. They applied GAN algorithm discriminator to identify the similarity between the generated output and the requested output of an urban pattern map. The algorithmic model was tested using new input map images from outside the training dataset. Visual comparisons were then conducted between the urban patterns and structure of the test output and real images of urban maps. Evaluation criteria included continuity of existing roads and proposals of new ones, identification of street network topology, and imitating the building morphology and typology [8].
Zhou and Park applied ML to generate stylistic alternative images from an input dataset composed of sketch-like drawings and textual attributes labels such as architect name, building type, year of construction, and size and location of project. The training dataset was obtained from real design images of well-known architects such as Foster and Zaha Hadid [4].
Deshpande et al. applied ML to analyze and simulate daylight performance of office building parametric design models. The training input is a synthetic dataset generated using a parametric model applied in Rhino, Grasshopper, and LadyBug tools to derive the daylight data of DA and sDA for different building configurations such as H and L shape typologies. After training, the ML model allows real-time prediction of 3D building daylight [9].
Cantemir and Kandemir applied CNN, as a deep learning algorithmic technique to analyze images and recognize architectural elements from different styles by classifying building patterns based on their façade into Gothic, Modernism, and Deconstruction styles. ML dataset included 1,043 images divided into training, testing and validation datasets. 835 images of training dataset were used to learn the relationships between them as input and the architectural styles as output [5].
Based on the above, there are different ways to approach machine learning in architecture. Machine learning algorithms can translate design parameter inputs such as building regulations or climate conditions into a building layout. Additionally, it is a decision-making process to help the architect evaluate alternative solutions and improve design accuracy and efficiency. ML algorithms such as GAN and CNN have been implemented in architecture as image-based ML to recognize the pattern of data and explore the design style. In fact, current machine learning experiments demonstrate the ability to handle specific design disciplines such as architectural design, urban planning, and environmental design. They consider different design features: formal, functional, and performance aspects that are represented using techniques such as images, graphs, drawings in 2D or 3D, text, etc. In addition, machine learning algorithms can be applied to diverse training datasets for the same or different tasks. Therefore, the paper focuses on an in-depth examination of both the design tasks that machine learning performs, and the training dataset required to accomplish the tasks.
3 Research aims and methodology
This paper aims to develop a comprehensive understanding of the potential of machine learning in architectural design. It seeks to identify the design tasks performed using machine learning techniques in architecture and to describe the design datasets needed to train machine learning algorithms to perform the required tasks.
The paper achieved this aim by adopting an inductive conceptual analysis of recent experiments on integrating ML in architecture research. The investigation had concentrated on CUMINCAD (Fig. 1) as a specialist open access index platform for Computer Aided Architectural Design CAAD. Most papers are published in specialist conferences (eCAADe, CAADRIA, ACADIA, CAAD futures, SIGraDi, and ASCAAD). The platform has 123 papers, which have “Machine Learning” in their titles. The search focused on the most recent articles published between 2019 and 2024 that implement machine learning in the field of architectural and urban design. According to that, the investigated papers were 51. The results of thematic analysis of design tasks and training datasets are summarized in Tables 1 and 2.
CUMINCAD index platform for open access CAAD publications [10]
Citation: International Review of Applied Sciences and Engineering 2025; 10.1556/1848.2025.00943
4 Machine learning for architectural design: tasks and datasets
Previous studies implemented ML to satisfy different design tasks using varied types of training dataset. To reach a comprehensive vision of ML experimentation in architectural design, the paper investigated 51 papers to determine the following aspects:
- -Types of design tasks achieved by ML.
- -Specifications of training dataset required to learn how to achieve the task.
The introduction of ML into architectural design practices has focused on three main tasks, they are generating design, predicting performance, recognizing and categorizing design space. Table 1 shows the tasks concluded from the 51 investigated studies.
Machine learning tasks concluded from the investigated studies
| Machine learning tasks | |
| Design generation | Generate brick patterns [11] |
| Generate conceptual shapes [12] | |
| Generate 3D spatial configurations of buildings [13] | |
| Generate new images with intervening of style images [14] | |
| Predict the road networks distribution on new urban layouts [15] | |
| Generate 2D image from semantic representation of the form of design [16] | |
| Predict building sections to conclude the missing parts [17] | |
| Generate architectural forms using strategy of blending of different styles [18] | |
| Generate morphological elements for the vacant waterfront [19] | |
| Generate 3D design from a graph representation [7] | |
| Produce new 3D parametric geometries from multiple object scripts [20] | |
| Create urban plans to assist the architect in generating building configuration details within a specific city condition [21] | |
| Generate new designs for timber joints [22] | |
| Generate layout plan of garden for specific site conditions [23] | |
| Restore designs of traditional Chinese Garden [24] | |
| Generate design form of urban fabric based on the city's structure [25] | |
| Design performance | Applying fire codes and regulations [26] |
| Predict residents' health status based on urban conditions [27] | |
| Predict building lighting measures: DA & sDA [9] | |
| Predicting the thermal efficiency of buildings and determining the parameters affecting the design [28] | |
| Predict the Global Warming Potential (GWP) based on a building massing, typology and location [29] | |
| Predict the surrounding conditions impacts on urban blocks [30] | |
| Predict the impact of different solutions on thermo-hygrometric conditions in heritage building [31] | |
| Exploring the relationships between dynamic facade design parameters and interior performance, including Energy Use Intensity (EUI) and Daylight Glare Probability (DGP) [32] | |
| Predict daylight performance in a building design [33, 34] | |
| Predict cost benefit and thermal performance [35] | |
| Evaluation of three common performance metrics: site energy, illuminance, and both [36] | |
| Measure performance indicators of landscape development plans in new design [37] | |
| Evaluate the environmental metrics of facade systems including sunlight hours and radiation values [20] | |
| Predict daylight provision and levels of overheating in dwellings [38] | |
| Optimize building facades such as window width and g-value to achieve optimal clustering of windows for desired level of daylight provision [39] | |
| Improve indoor environment performance and comfort [40] | |
| Assess Window-to-Wall Ratios [41] | |
| Predicting the factors of building natural lighting, energy use, and other performance metrics [42] | |
| Design recognition and classification | Recognition of clash detection [43] |
| Clustering the design precedents into style groups [44] | |
| Predict the analysis of Spatial Openness Index (SOI) for large scale urban fabrics [45] | |
| Recognizing and labelling rooms of residential apartments [46] | |
| Clustering villages based on spatial features [47] | |
| Identify functional and typological traits of building plans [48] | |
| Feature extraction and cluster analysis of Detroit city [49] | |
| Classification model for predicting construction type based on massing model and target GWP [29] | |
| Identify different types of land use from raw satellite Imagery [50] | |
| Image classifier tool to identify the rooms that usually compound a dwelling [36] | |
| Predicting specific properties of reusing components from existing buildings [51] | |
| Classify the data of screen cluster, considering both the urban morphology and the properties of screen buildings [52] | |
| Identify clusters of identical urban conditions to produce a picture of the morphological structure of the city [53] | |
| Reveal five types of traditional patterns of Chinese village [54] | |
| Classify the Chinese courtyard housing type variables from the perspective of visual-spatial perception [55] | |
| Predict a target plot land use built on its surrounding environment [56] | |
| Classify functions of room in diagrams of residential building 2D plan [57] | |
| Predict the design of bedroom plan using labelled drawings [58] | |
The application of ML into architecture practices reveals that architectural design may rely on ML techniques to take advantage of data-rich computational environments and workflows [59]. The ML algorithmic model can learn physical or conceptual patterns of architectural design elements that are represented computationally. Table 2 shows the specification of design datasets concluded from the 51 investigated studies.
Machine learning training dataset concluded from the investigated studies
| Source and type of training dataset | |
| Source of training dataset | clash detection reports [43] |
| project case studies [26] | |
| Fire code and regulation [26] | |
| architectural precedents [44] | |
| brick patterns for varied wall geometries [11] | |
| 3D designs based on existing buildings [12] | |
| height maps and SOI maps of urban fabrics [45] | |
| 2D floor plan datasets constructed from scratch [13] | |
| 454 images of floor plans from previous projects [46] | |
| POI data from Open Street Map and Local Data for Better Health [27] | |
| 3D model of an existing projects [14] | |
| geographical information data in axial maps of villages [47] | |
| synthetic data derived from environmental simulation of the parametric model [9] | |
| Three case studies of housing estates [15] | |
| Plans of religious buildings [48] | |
| Abstracted description of the design brief [16] | |
| Images of 80 sections [18] | |
| Images of 2D section and final 3D model of famous buildings [18] | |
| Urban projects had successful regeneration [19] | |
| Collected information about Detroit city [49] | |
| Granary buildings [28] | |
| Online platform of architectural design competition [7] | |
| Open-source data including GWP, the type of construction, the building size of (GFA, FC), and its location [29] | |
| 500 solar radiation simulations for each case of tow small-scale urban blocks [30] | |
| synthetic dataset for energy simulations of cultural heritage museum building [31] | |
| Light and energy simulation data of four models of dynamic facade with varied parameters of unit length, width, patterns, etc [32] | |
| 500 simulation cases of parametric prototype model [33] | |
| Parametric design alternatives for a base case design scenario [34] | |
| 350 United States airports [50] | |
| Synthetic datasets using parametric model to generate different roof and wall coverage [35] | |
| Houses and apartments in Lisbon available in the search engines of real estate agency [36] | |
| Residential and commercial mid-rise tower project [36] | |
| 3749 evaluations of performance were done by 36 professionals to examine six criteria of sustainability in 32 neighborhoods' designs [37] | |
| Residential buildings, subject to renovation or demolition, built between 1925–1975 in Barcelona and Zurich [51] | |
| Building data was concentrated on screen clusters from Shanghai and Chongqing [52] | |
| Building footprints and heights, parcels geometries, neighborhood boundaries, and streets [53] | |
| Randomized input of parametric script data. [20] | |
| Multiple data sets of parametric geometrical objects. [20] | |
| Design parameters of existing high-rise building represent different combinations of the window-to-floor value, g-value, visual transmittance of the glazing system, etc. [38] | |
| Parametric model represents room orientation and dimensions, window dimensions, and the angle of obstruction [39] | |
| the data from different cities [21] | |
| 2,187 data samples of louvre shades and film shades [40] | |
| buildings in the Bronx, New York [41] | |
| 6,819 traditional Chinese villages [54] | |
| Beijing Siheyuan plans using samples from an ancient map [55] | |
| Nanjing, Jiangsu, China, as a case study and its old town as the study area [56] | |
| Synthetic data generated automatically using parametric modelling software to [42] | |
| Parametric timber joints [22] | |
| Classic cases of traditional Chinese private gardens [23] | |
| Traditional Chinese Garden dataset [24] | |
| A sample of 1,586 residential rooms from 165 building plans. The rooms are classified into nine types [57] | |
| 179 bedroom plan drawings collected from lianjia.com [58] | |
| A sample of 136 European cities according to the rules of Nolli map [25] | |
| Representation types of training dataset | Image of report [43] |
| Text of fire code [26] | |
| Revit drawing [26] | |
| Numerical data form [44] | |
| Input image of wall boundary and output image of brick wall [11] | |
| Input text to output image or input image to output 3D shapes [12] | |
| Map images using colour information [45] | |
| Graph-based 3D geometry [13] | |
| Floor plans in CAD drawings, hand drawing and marketing images [46] | |
| Maps of streets [27] | |
| 2D style transferred images and 3D geometries [14] | |
| Images of axial maps [47] | |
| A parametric model and environmental simulation [9] | |
| Satellite images of housing estates [15] | |
| Building plans [48] | |
| Texts and images [16] | |
| Section images of interior space drawing and exterior shell drawing [17] | |
| 2D section images and 3D model [17] | |
| Image pairs before-and-after urban regeneration [19] | |
| Image of Detroit city [49] | |
| Parameters such as building height, ratio, overall scale, outdoor temperature, orientation, coefficient of the wall and roof heat transmission [28] | |
| Graph and subgraphs representations [7] | |
| Benchmarks for Global Warming Potential and building parameters of location, type, floor count, and gross floor area [29] | |
| Regulations and solar radiation values of two sites [30] | |
| Measurement data of the cultural heritage museum environments, real sensor network data and synthetic energy simulations data [31] | |
| Parameters of facade and performance metrics including the values of “Energy Use Intensity” and “Daylight Glare Probability” [32] | |
| Facade parameters and light measures (DLA, UDI, DGP) [33] | |
| Data of design changes and the resulting daylight levels [34] | |
| Satellite images of airports and land use parameters around the airport using computer vision mapped against passenger footfall numbers [50] | |
| Images of roof and wall panel configurations onto the base façade [35] | |
| Images of dwellings and images of four room ID categories are Bathroom, Bedroom, Kitchen, and Living room [36] | |
| training sets, based on actual simulations [36] | |
| six sustainability criteria [37] | |
| image dataset of residential building [51] | |
| Diagrammatic Images, morphological maps of urban screen clusters and screen-based building features [52] | |
| Diagrammatic images represent the urban context of each building in the city [53] | |
| calculations such as sunlight hours and radiation values [20] | |
| parametric design dataset of types of object and transformation rules [20] | |
| Values of DF and DH and different design parameters: ratio of window to floor, obstruction level, g-value, and glazing visible transmittance [38] | |
| Parameters of the model are: room orientation and dimensions, dimensions of window, and angle of obstruction [39] | |
| Image pairs of the site conditions including roads, green areas, rivers and the buildings configuration [21] | |
| Parameters of adjustable shading of façade and the environmental performance of indoor [40] | |
| Window to wall ratios (WWRs) [41] | |
| Village patterns [54] | |
| Plans and visual graph analysis (VGA) [55] | |
| Land use grid map [56] | |
| Design performance metrics [42] | |
| Graph format [22] | |
| Image-to-image translation of site boundary of garden to output layout plan of garden [23] | |
| Paired text descriptions and images [24] | |
| Image of plan with the basic elements such walls, doors, windows and room boundaries [57] | |
| Vectorized data and image data [58] | |
| Image-to-Image of urban fabric map [25] | |
Built on the induced data in Tables 1 and 2, the framework for applying ML to architectural design problems is established to elaborate the definition of design tasks and the training dataset.
4.1 Types of design tasks in machine learning experiments
Table 1 reveals that many design tasks underline the use of ML in architectural design. It can be used in the design generation process, the design optimization process, and the design recognition and classification process.
4.1.1 Using machine learning for design generation
In architectural design, ML tasks are moving from working as analysis tools of design datasets towards new functions focused on design synthesis and creation [59]. ML-aided design supports the data-driven derivation of various design possibilities. It has remarkable capabilities in the generation of innovative 2D and 3D architectural designs such as floor plans, facades and building masses [60]. Zhang and Huang [18] considered the strategy of “style blending” using ML is innovative design method of form finding. It generates new architectural forms based on merging existing forms including “2D drawings and 3D models” to derive new forms from varied architectural styles.
Previous studies in Table 1 examined the efficacy of using ML as a generative design method in architecture and urban design experiments. This paper identified five main scenarios of design generation using ML depending on their training datasets which can be pre-design data, conceptual design data, parametric design data, typological design date, and restorative design data as clarified in Fig. 2.
Types of learning dataset used in ML generation process
Citation: International Review of Applied Sciences and Engineering 2025; 10.1556/1848.2025.00943
Firstly: the design generation is based on pre-design data, such as design ideas, requirements, constraints, rules, and descriptions. Wu [1] implemented ML model in the generation of building footprints in varied urban fabrics of Boston City. The generation process covered the housing floorplan fully or partially. Huang and Zheng [19] applied a “Generative Adversarial Network (GAN)” as a machine-learning technique to generate a new design of an industrial waterfront based on the design characteristics and morphological elements of an empty industrial waterfront. The generation process included subdivisions of empty land and deciding land use. Boim et al. [8] applied the GAN model also as an experimental tool for learning the design datasets of existing urban morphological patterns and reproducing new spatial patterns for non-planned settlements. The generative model learned from local patterns to generate new multiple urban design alternatives with similar morphology to the original urban dataset by adjusting existing urban design fabric. The model can complete the pattern of the partial map in a similar way to the existing morphology. del Campo [61] applied ML in the automated generation of images based on design descriptions written in natural language during the initial design process.
Secondly: The design generation is based on conceptual design data, such as plan layout, design sketch, etc. ML can convert 2D conceptual sketch images into a 3D design model, such as transferring style images into building massing [62]. In addition, ML used sketch and textual data input to generate a predicted output image. Applying supervised learning from design precedents can generate varied floor plans based on the changes in their program [4]. Andreou et al. [63] explored how ML can be implemented to generate conceptual shapes by transforming input text dataset to image or converting image to three dimensional shapes. Karabagli et al. [7] implemented ML in the conceptual design process to convert a graph representation as a design input into a 3D massing model. Translating the training dataset of graph representations to floorplan designs was done via the “HouseGAN++” model and then generating 3D volumetric design output using Blender3D. Zhou and Park [4] applied ML in forming stylistic variations of the initial sketch as user input. The training dataset included 10,000 pairs of input and output design images gathered from the websites of well-known architects such as Foster and Partners and Zaha Hadid. The “XDog” application can generate sketch-like drawings, similar to human sketches, to be paired with the collected images. Text features such as the architect's name, building type, year of construction, size, and location were added to label each collected image. Pairs of generated sketch images as input dataset and the website images as output dataset were used to train the ML.
Thirdly: The design generation is based on parametric design data. ML and generative processes of parametric design enable the production of alternative design variants [59]. It enables the designer to combine different parametric design components to generate new complex geometry. Sebestyen et al. [20] used ML to derive multiple geometric models of design objects based on a parametric design dataset. The parametric dataset of each design input involves one of the basic types of design object, such as a cube, and one of the transformation rules of the object geometry. The parametric values of the sample of generated object geometry have been randomly chosen. Zheng and Yuan [2] applied an artificial neural network to learn different design parameters of building features. They combined input parameters of two design styles to generate tower geometry. Bauscher et al. [13] used ML to generate 3D parametric geometries of architectural style based on machine-readable input dataset of graph representation. ML model derives a new design output based on learned logic which combines varied parametric design features of input dataset.
Fourthly: The design generation is based on typological design data. ML algorithms support the derivation of typological design. In this case, the training process is based on a dataset composed of a corpus of object type as a group of urban spaces or buildings having similar essential characteristics. For example, GAN algorithms use labelling to identify the shared spatial elements among the designs of private gardens in China and to apply them in the generation of a new layout plan [23].
Fifthly: The design generation is based on historical restoration data. ML plays an important role in the discipline of architectural conservation of heritage buildings. Its techniques can be used to support architects and guide the decisions on historic and ruined architecture restoration. In the learning process of neural network algorithms such as CNN and GAN, the style, aesthetic features, and layout of traditional and historical design can be captured to be applied in the restoration of a new design or part of existing design [5, 24].
Machine learning algorithms such as deep learning have the ability to recognize design features and patterns by analyzing massive data sets, facilitating the recovery of damaged or missing elements. They can be implemented to interpret documented datasets and to assess the missing historical elements, textures, patterns, or colors, and to detect their typologies [64]. For example, Karadag applied ML to interpret the documented dataset and to regenerate the missing parts of ancient buildings. He implemented GAN model to predict the basic geometry of damaged components of historical architectural details of Ottoman tombs [64].
4.1.2 Using machine learning for design performance analysis
Throughout the last two decades, ML has been widely implemented in the analysis and prediction of architectural design performance. Introducing ML training datasets with performance evaluations such as solar radiation, structural performance, etc., supports quick decision-making in the early stage of the design process. This activity replaces the simultaneous simulation of a current design performance with a machine-learning predictive model learned from a dataset of analytical simulation results [6]. ML model estimates the mapping relations between the input and output datasets to predict the objective function for the design optimization.
ML can advance the performative instrumentalization of parametric modeling. On one hand, the performance metric can be prescribed to guide the design generation process resulting in varied outputs. The ML algorithm translates the input dataset of performance parameters into the output dataset of the building layout [59]. For example, ML can predict the effects of different thermo-active material patterns on the resulting deformation of the panel at a particular temperature [6]. On the other hand, ML can apply supervised learning to predict the design simulation results by learning the pattern of input-output datasets [59]. In both cases, using minimum effort, ML allows designers to make automated and quick approximate predictions of design performance and reach design recommendations regarding massing options or vice versa [9]. Therefore, incorporating ML into the early design phase can accelerate environmental performance prediction when changing design parameters is possible [33].
Obeidat et al. [67] applied supervised algorithms using SVM and ANN to predict energy usage of parametric residential units at the early design stage. They used Grasshopper plugin as a parametric generative tool and honeybee plugin for energy consumption calculations. Raanan et al. [37] trained neural network algorithms to imitate the expert evaluations of sustainability properties of neighborhood and landscape plans. Deshpande et al. [9] used ML in real-time analyses of environmental performance during conceptual design. They trained the ML system using an automatic synthetic dataset generated via the environmental simulation of a parametric design model to identify optimal building size, window-to-wall ratio of façade, and orientation. The aim is to predict and maximize building daylight measures such as “Daylight Autonomy” and “Spatial Daylight Autonomy” and to reduce the costs and time of building simulation. The parametric model of office building in Singapore as a case study was applied in Rhino, Grasshopper, and LadyBug software to automate the generation of daylight datasets for different configurations and locations, as clarified in Fig. 3.
Daylight optimization using ML [9]
Citation: International Review of Applied Sciences and Engineering 2025; 10.1556/1848.2025.00943
Li et al. [33] implemented ML for daylight imitation of the adaptive facades. They investigated how different parameters of adaptive facade affect indoor daylighting. They used a ML model to imitate the simulation of light and predict the effects of parametric adaptive facade on light measures such as “Daylight Illuminance”, “Daylight Glare Probability”, and “Daylight Autonomy”. The adaptive facades in this study have modular units: rectangular, hexagonal, and triangular arranged in vertical and horizontal arrays to allow natural light admission to the room, as clarified in Fig. 4.
Parametric training dataset for façade design [33]
Citation: International Review of Applied Sciences and Engineering 2025; 10.1556/1848.2025.00943
The training dataset of the artificial neural network algorithmic model included images, text, etc., to represent 500 design cases for each parametric façade model and the values of their three light indicators under weather data. Design parameters included the unit width and length, the distance of the unit, transparency of material, diverse patterns, etc. [33].
Furthermore, ML models are used also to assess the performance of finishing in historical buildings. Deep learning technology can help provide useful and semantic information of digital heritage by recognizing different building elements, surface circumstance and cracks with details [65]. For example, Karimi et al. presented an automated model for tile damage detection in Portuguese heritage buildings, with the aim of assisting visual inspections using the automatic tile damage detection [66], and to anticipate potential cracking or discoloration to take immediate preventive action [65].
4.1.3 Using machine learning for design recognition
ML plays an important role in design recognition using classification or clustering processes.
- -Design Recognition via Classification
The classification task assigns predefined classes to a dataset of design objects based on their features. It is a supervised learning problem that learns from the pairs of input-output training datasets to predict the class of a new object. The classification task of ML predictive model is to map an object to a class based on a training dataset of input-output pairs.
The frequent application of artificial neural networks as a ML technique in image recognition is done by detecting features over images firstly, classifying these features, and distinguishing between them secondly. The neural networks have varied applications in architectural design, such as classifying spaces, categorizing architectural drawings, differentiating between architectural plans and sections, distinguishing between designs of different architects, and making distinctions among different building typologies based on floor plans [60, 23]. Furthermore, deep learning is used to classify objects of cultural heritage based on three-dimensional point clouds [65].
The recurrent use of artificial neural networks is the classification of designs according to their typologies, for example, identifying Palladian villa designs from thousands of villa images. During the training cycle, the system specifies the characters of Palladian architecture autonomously without any explicit definition of these characteristics. After training, ML can quickly and accurately infer if the image satisfies the design criteria of the Palladian style [6].
ML can distinguish the details of different building styles. Cantemir and Kandemir [5] applied neural network techniques in classifying building patterns based on training datasets of architectural façade styles such as Gothics, Modernism, and Deconstruction. In addition, this task plays a vital role in architectural conservation. The neural network algorithms can learn the aesthetic characteristics of historical buildings to guide and direct restoration decisions.
- -Design Recognition via Clustering
The clustering task determines undefined classes by discovering the similarities and differences in the underlying features of the design dataset. It identifies similar properties among design objects and groups them according to their identical features, distinguishing the group from others [24]. It is an unsupervised learning activity where mappings of input-to-output are not required. The training dataset contains only the training inputs where the output is unknown [71]. Using ML in clustering datasets of architectural designs incorporates the design analyses, comparison, and prediction of their similar morphological characters and elements [3]. Therefore, this task can be used for novelty detection [24].
Alymani et al. adopt ML experiment to categorize input dataset of 900 topological graph of 3D building prototypes based on their ground relationships [24]. Tochaiwat and Seniwong [69] used ML for clustering housing projects based on land price, local parks, etc. They implemented ML techniques in categorizing 179 housing projects in Thailand, where monthly sales rates were above the average. The characteristics of housing projects that contributed to the determination of housing clusters were land price, local park, and a bus stop.
Kim and Huang [70] applied ML to recognize the formal similarity among 3D volume characteristics of school designs. They considered that form clustering supports imagining the relations among associated forms in the dataset that exceeds human eye recognition. They examined the abilities of this technique to learn and understand the correlated features of 3D architectural design and conducted a comparative analysis of their formal characteristics. They applied ML in clustering the dataset of more than 360 architectural projects designed for 63 school competitions between 2009 and 2021 in Switzerland. ML explored and learned the distinctive formal characteristics of the 3D volume design dataset represented in “point clouds” format and then clustered different groups having similarities among their forms.
Millan et al. [71] applied ML in tracking the initial design process to gain a better understanding of design intentions. Tracing the design process aims to realize best design practices by recognizing the sequence of design decisions and actions. They used ML for clustering snapshots of design shelters having similar typologies or problem-solving strategies. Training datasets included fifty-two snapshots of shelter designs to be analyzed to extract the design knowledge, identify the design strategies, and cluster design solutions having similar typologies or problem-solving strategies.
4.2 Specifications of training dataset in machine learning
Training data is the cornerstone of the success or failure of a ML algorithmic model. This paper seeks to explore, investigate, and reveal the different types of datasets used in ML training and testing that participate in solving architectural design problems. Specification of ML datasets applied in architectural design can be defined in terms of the data sources and the descriptions of training data, as follows:
- -Source of training datasets.
- -Descriptions of training datasets are defined in terms of:
Representation techniques of the training dataset
Relations between input and output data.
4.2.1 Sources of training datasets
There are two types of datasets in ML: original data and synthesized data. The original dataset exists in reality designs, drawings, sketches, etc. The synthetic dataset is artificially composed via computer-based automated generation such as parametric modeling [6, 9]. Parametric generative processes can produce an enormous dataset of all possible design alternatives [59]. The design space size in the parametric modeling depends on the number of parametric variables in the model [20].
Examples of real training datasets are the scenarios of regenerating industrial waterfront projects including the maps of the land-use and street network before and after design collected from satellite maps and the New York City data platform [19], also traditional knowledge of real datasets of local urban planning practices [8], existing neighborhood plans and landscape with their sustainability evaluations assessed by professional experts [37], housing projects in Thailand [69], architectural design competition projects [7, 70], and academic designs of shelter done by students [71], a dataset of 10,000 internet images of architectural designs [4], real projects of “24 Highschool Shenzhen” [61], photos of newly constructed projects during (2009–2019) in Hangzhou [3], footages of five iconic buildings in downtown Los Angeles [62]. In addition, two hundreds plan drawings of historical buildings were used as a dataset for training ML [64].
On the other hand, Tsigkari et al. applied parametric generation and analytical simulation of thousands of synthetic office building floor plans with their visual and spatial analyses [6]. In addition, Zheng and Yuan used synthetic training datasets composed of 900 randomly generated forms of existing building styles [2]. Bauscher et al. constructed ML training dataset of hundred designs from four modernist well-known buildings. The training dataset was generated from parametric variation of building elements without losing the spatial concept [13]. In addition, Alymani et al. generated synthetic training dataset of geometrical models from varied topological ground relationships [68].
4.2.2 Descriptions of training datasets
Training processes in ML rely on different types of design inputs. The input dataset is the key to ML success. Previous studies considered three conditions in defining their training datasets to obtain reliable outputs. They are:
Firstly, the quantity of the input dataset is a crucial factor. ML algorithms need a large amount of design datasets to be able to learn the model and predict design outcomes [4].
Secondly, the quality of the input dataset is also a key factor. The training datasets must be high-quality designs and match recognized patterns for the algorithmic model to learn [6].
Thirdly, the architectural features and relationships of the training dataset that ML algorithms can learn must be in a computationally processable format such as a drawing, graph, image, text, etc.
The paper describes the input and output of training and testing datasets in ML. It focuses on their representation techniques and the relations between design input and output.
- -Representation techniques of training and testing dataset
ML has computational techniques to extract design knowledge from the training dataset. However, there is a need to pre-process design data before inputting it into the ML algorithmic model. There are specific types of design representations that enable knowledge extraction in ML. Previous experiments applied visual and (or) textual representation techniques. Numerous experiments in ML rely on “2D pixel” image recognition [24]. Building components can be recognized using dataset of “3D point cloud data” [64]. In addition, image segmentation techniques using deep learning have revolutionized the way heritage experts recognize and classify components of historic buildings [65].
Zhou and Park represented training data from 10,000 input images using freehand sketches, line drawings, and text descriptions such as architect name, project size, building category, year of construction, and location. They represented training data for design output using architectural-style-colored images [4]. In addition, Raanan et al. used a dataset of 625 input images for training and 155 input images for testing. The training on the output dataset was based on qualitative assessments by 36 experts, including 3,749 performance judgments of 32 neighborhood designs considering six criteria of sustainability [37]. Brown et al. [46] defined design features of the floor plan that enable the training process in Convolutional Neural Networks CNNs. For example, they used plan lines to define walls and boundaries, patterns to define plan layout, and rules of relational data regarding room size, layout, and shape. In addition, they added text, icons, and symbols to represent room functions, such as bedroom text and a symbol of bed [23].
Millan et al. defined a training dataset using six snapshots of shelter projects and represented each snapshot in terms of nine textual variables [71]. Cantemir and Kandemir used images from 40 epochs as a training dataset [5]. Özel implemented sketches of artistic stylized images for landmark buildings as the input to the training dataset and their 3D model image as the output to the training dataset [62]. Boim et al. depended on a training dataset including pairs of input partially blank maps and output complete maps. The urban pattern training datasets had 220 JPEG maps with building masses and roads as a sufficient training format for Generative Adversarial Network GAN [8]. Huang and Zheng represented design features of industrial waterfront maps using linear shapes, irregular boundaries, urban networks, and labels of colored land use [19].
Karabagli et al. applied an input training dataset using graphs as machine-readable representations and an output training dataset using 3D functional massing models [7]. Bauscher et al. considered that using graph representations (elements and relationships) in ML model store more information and date than image representations [13]. Also, Alymani et al. represented the input dataset of buildings' geometrical models and their contextual relations into topological architectural graphs [68].
Deshpande et al. defined the training dataset of office building drawings using parameters such as floorplan arrangement, length, width, height, window-to-wall ratio in cardinal directions, location, and rotation [9]. Li et al. represented a training dataset of parametric facades using input features such as length, width, and distance of a unit, units' patterns, transparency of the material, etc. The output dataset of indoor lighting used for training included numerical results of Daylight Illuminance, Daylight Glare Probability, and Daylight Autonomy [33].
Tochaiwat and Seniwong [69] used market reports to represent a training dataset of 179 housing projects. del Campo used language in written form to describe the building program and activities as design inputs for the training dataset of the “AttnGAN Algorithm” [61]. Xia et al. applied a training dataset of residential buildings using morphological elements such as material type, color, balcony type, balcony shape, window type, sunshade, etc., and factors of site economics such as building height, location, price, construction time, etc [3].
Other studies applied special visual representation techniques. Kim and Huang used input datasets represented in point clouds (Fig. 5) as computational learnable formats to translate NURBS-based 3D massing models of 366 architectural forms [70]. Sebestyen et al. applied data points of a 3D grid size 32 × 32 × 32 to represent the parametric geometry of input and output datasets with differing dimensions [20]. Zheng and Yuan implemented a grid of 3D points (x, y, z) to describe NURBS and represent the parametric features of 900 tower-like geometries as a training dataset [2].
- -Relations between input and output data
In a data-driven approach, making decisions relies on collected data instead of intuition. This method sets up relationships between the input dataset and the output dataset. Built on previous experiments on using ML in architecture design, this paper identified different types of relations between the pairs of training input-output data. There are types of paired representation used for input-output training datasets. For example, they can be:
Image-to-Image is a training model that takes an image as design inputs and reconstructs a new version of this image as a design output. For example, in the design generation process, ML algorithm can translate a color map as a design input into an output facade image.
Text-to-Text is a training model that uses textual descriptions as both design input and output. For example, the input dataset can include building descriptions such as functions, floor areas, number of floors, materials, etc. and output datasets is the calculation of cost.
Text-to-Image is a training model that uses textual descriptions as design inputs and then generates visual representations as design output. For example, ML dataset for traditional Chinese garden has paired descriptions of text and images. Learning the model is done by mapping from the textual description of a sample of garden designs to the actual images of these garden designs [24]. del Campo applied ML in the automated production of images in response to input descriptions using natural language [61].
Image-to-Text is a training model that uses visual representations as design input and then generates textual descriptions as design output.
Graph-to-Image is a training model that converts input descriptions of graphs composed of nodes, edges and vectors into visual representations of image output.
In addition, machine learning relies on a set of operations to transform a design input dataset into a design output dataset, such as refinement, embodiment, sorting, or evaluation that operate on the training dataset inputs. The different processes are clarified as follows:
Using point cloud to dataset representation [70]
Citation: International Review of Applied Sciences and Engineering 2025; 10.1556/1848.2025.00943
4.2.2.1 Design output is the refinement of design input
Design refinement is a process of making improvements by adding details, descriptions, etc. Wu applied the ML algorithm Pix2Pix to translate the input image of rectangular-colored shapes to output a photorealistic more detailed image, as shown in Fig. 6 [1].
Training dataset: Input and output images [1]
Citation: International Review of Applied Sciences and Engineering 2025; 10.1556/1848.2025.00943
Huang and Zheng trained a ML neural network using the image-to-image method to identify the morphological features on the masterplan scale of industrial land and to transform the input real industrial waterfront map into a predicted output map with land-use, building footprint, and street network as clarified in Fig. 7 [19].
Using map images for input and output training datasets [19]
Citation: International Review of Applied Sciences and Engineering 2025; 10.1556/1848.2025.00943
Boim et al. taught ML using 200 pairs of input datasets composed of a partially blank map (by erasing the right part) and an output dataset including a complete map. They used the ML model to refine existing maps by keeping existing roads, suggesting new roads, and reproducing similar building morphology and typology [8]. Brown et al. [46] applied ML to refine the input dataset of architectural drawings of floor plans by automatically labeling six classes of rooms (bedroom, living room, kitchen, dining room, bathroom, and balcony).
4.2.2.2 Design output is the embodiment of design input
Embodiment is the process of giving a form to non-physical sketches and ideas. Karabagli et al. used ML to transform input datasets composed of graph representations into output datasets embodied as massing models [7]. Sebestyen et al. trained a ML model on the design parameters as an input dataset to be embodied in a 3D shape output dataset [20]. Zhou and Park used ML to transform the input dataset of sketches and descriptive texts into an output dataset represented as a stylistic image [4]. del Campo converted natural language as input descriptions to a visual image output [61]. Özel implemented ML to transform 2D conceptual image inputs into 3D model outputs [62]. Zheng and Yuan applied ANN to convert design parameters of tower-like form as input data into the building geometry as output data [2]. Wu trained the neural network using over one thousand pairs of input and predicted output images. The input images on the left of Fig. 8 include design constraints such as roads, surrounding buildings, orientation, and location. The output right images in Fig. 8 locate a house floorplan on the site [1].
Training dataset: Input and output images [1]
Citation: International Review of Applied Sciences and Engineering 2025; 10.1556/1848.2025.00943
4.2.2.3 Design output is the sorting of design input
Sorting is the process of separating input dataset designs into groups. This process requires the abstraction of design inputs where the design outputs are familiar design features derived from classifying or clustering the input dataset. Tochaiwat and Seniwong applied ML to sort housing projects as design input to predicted classes as output data based on the proximity to services such as a main road, train station, bus stop, hospital and department store [69]. Kim and Huang trained ML to sort the input dataset of 3D design projects into output clusters based on details such as courtyards, voids, facades, and roof surface complexity [70]. Xia et al. implemented ANN to sort input datasets of residential buildings into recognized classes based on their morphological elements such as material type, color, balcony type, balcony shape, window type, and sunshade, and their site economic factors such as building height, location, price, construction time [3]. Millan et al. applied the ML model to sort shelter designs as design input into output clusters based on similar typologies or problem-solving strategies [71].
4.2.2.4 Design output is the evaluation result of design input
Evaluation is the process of making judgments regarding the performance of design input. Tsigkari et al. applied ML to analyze thousands of floor plans as design inputs and to obtain output results for their visual and spatial properties [6]. Deshpande et al. implemented a parametric model of simplified building as an input dataset in ANN to predict building daylight metrics: Daylight Autonomy and spatial Daylight Autonomy as design outputs [9]. Li et al. built a machine-learning model based on parametric adaptive façade as input data for predicting light simulation criteria: Daylight Illuminance, Daylight Glare Probability, and Daylight Autonomy as output data [33]. Raanan et al. used the ML model to evaluate landscape designs as design input based on sustainability criteria as design output [37].
5 Discussion and conclusions
Machine Learning tools are applied as efficient analytical tools in processing and exploring design data. A review of recent research efforts on machine-learning experimentation in architecture makes clear that machine-learning methods are beginning to find their way into the architecture and construction industry.
This paper seeks to provide a comprehensive understanding of ML implementation in architectural design. It presented thematic research on using ML in architectural and urban design. ML methods enable digital tools to learn from data the self-execution of specific design tasks and the capability to solve some design problems. According to that, questions are raised about what tasks ML can perform on the one hand; and how the ML input dataset is defined, represented, and transformed into output predictions on the other hand.
Table 3 clarifies design tasks concluded from recent ML experiments in architecture. ML was dedicated to design generation, optimization, and classification or clustering. The options for using ML as a generative tool depend on the type of training data, which can range from pre-design specifications, design idea sketches, parametric design values, typological design data, or restoration design data. A predictive ML model can calculate building performance by imitating computational simulation results or relevant experts' assessments, or it can identify building features based on prescribed performance. Design recognition via ML is applied to classify or cluster design datasets into predefined or non-predefined classes.
Design tasks using machine learning
| Machine Learning Design Tasks | Machine Learning for design generation | based on pre-design data |
| based on conceptual design data | ||
| based on parametric design data | ||
| Based on typological design data | ||
| based on historical restoration data | ||
| Machine Learning for design optimization | Optimization measures are prescribed as design input to derive the output design variables | |
| Optimization measures are calculated as output based on the input design variables | ||
| Machine Learning for design recognition | based on Classification into predefined classes | |
| based on Clustering into non-predefined classes |
ML decisions rely on the collected dataset. ML extracts significant insights from training datasets as the key to ML tasks' success. ML experiments in the architecture domain revealed diverse design data used in training and testing the algorithmic model. Table 4 shows the variation in training data regarding their sources, representation techniques, and input-output relationships.
Specifications of Machine Learning training dataset
| Specifications of training dataset | ||
| Sources of dataset | Real data | |
| Virtual data | ||
| Representation techniques of training dataset | Visual representation techniques | drawings |
| images | ||
| graphs | ||
| maps | ||
| sketches | ||
| others | ||
| Textual representation techniques | Natural language | |
| Script language | ||
| others | ||
| Relations between input and output data | Input-to-output representation | Image-to-Image |
| text-to-text | ||
| Image-to-text | ||
| text-to-Image | ||
| Image-to-Graph | ||
| Graph-to-Image | ||
| Graph-to-Graph | ||
| Drawings-to-Image | ||
| others | ||
| Input to output conversion processes | Refinement | |
| Embodiment | ||
| Sorting | ||
| Evaluation | ||
The ability of machine learning to handle different types of representations leads to the diversity of performed tasks. The use of visual techniques in training datasets such as 2D and 3D drawings, graphics, diagrams, and images activate the generative role of machine learning in architectural design. On the other hand, the use of texts such as numbers, equations, etc. enhances the role of performance prediction using ML in architectural design.
In conclusion, this study contributes to bridging the gap between theoretical potential and practical application of machine learning in architecture. It focuses, in particular, on the definition and representation of training datasets. It provides a comprehensive view of ML tasks in architectural design. The theoretical framework introduces some aspects not addressed in previous studies. It defines the roles of ML considering various training data and during different stages of the design process. In addition, this study elaborated the definition of training data in terms of its sources, how it is represented, and the relationship between its inputs and outputs.
Furthermore, this paper underscores the potential of machine learning to foster creativity, rather than merely automating routine tasks, emphasizing its transformative role in reshaping architectural innovation through digital technologies. This finding disagrees with the preconceived notion that using previous designs as training datasets makes ML models lacking innovative new designs; and is questionable. ML can contribute to creativity in the design process by supporting form finding. Generative ML, like other generative design methods such as shape grammars or parametric design, supports creativity resulting from generating multiple design solutions embodying a form-finding approach rather than focusing on a single design option that follows a traditional form-making approach. Design generation using ML allows both form-making if the training dataset is built on physical design features and form-finding if the training dataset depends on conceptual design characteristics.
Moreover, learning from preexisting design datasets enhances the designer's knowledge and insight. Recognizing design correlations via ML can track design strategy, problem-solving pathways, and design creativity by revealing hidden similarities among collected datasets of best projects. Also, when dealing with design styles, the computational abilities of ML can enhance innovation by recognizing and capturing the essence of architectural style that even critics and specialists may fail to perceive the human eye. Additionally, ML enables blending styles by combining different ones to make a new design.
ML can support design decisions using synthetic architectural datasets and automated, guided design processes. Incorporating routine ML tasks into the architectural design process such as performance prediction for varied parametric massing or plans, frees designers from doing routine tasks such as doing calculations. In this case, ML helps designers save time and effort on more creative tasks.
On the other hand, ML presents new opportunities for addressing critical challenges in sustainability, heritage restoration, and adaptive reuse, thereby supporting the evolution of these practices in a rapidly changing global context. Additionally, ML as a computer-aided design decision is capable to integrate with other computational design methods such as parametric design, genetic algorithms, building information modelling, etc.
Finally, despite the accelerating role that machine learning is playing in many fields, there are many challenges facing it. The accuracy of machine learning results depends on the reliability of the inputs. Therefore, it is necessary to ensure the validity of training dataset resources and avoid violating other designers' intellectual property rights. Also, using machine learning in heritage conservation has many challenges such as the availability of sufficient training dataset and the complex representation of varied point clouds for irregular sample of design dataset.
Future research should prioritize the integration of advanced machine learning methods, such as generative adversarial networks and reinforcement learning, to tackle complex architectural challenges, including dynamic energy modeling and the preservation of cultural heritage. Additionally, fostering greater interdisciplinary collaboration and creating high-quality open datasets would expand the applicability of these technologies, stimulate creativity, and promote innovation in architectural design and urban planning.
Nomenclature
| ANN | Artificial Neural Network |
| GAN | Generative Adversarial Network |
| CNN | Convolutional Neural Network |
| ML | Machine Learning |
| 2D | Two Dimensional |
| 3D | Three Dimensional |
| SVM | Support Vector Machine |
References
- [1]↑
C. Wu, “Machine learning in housing design: exploration of generative adversarial network in site plan / floorplan generation: Exploration of generative adversarial network in site plan / floorplan generation,” M. Arch. Thesis, Department of Architecture, Massachusetts Institute of Technology, 2020. https://dspace.mit.edu/handle/1721.1/129855.
- [2]↑
H. Zheng and P. F. Yuan, “A generative architectural and urban design method through artificial neural networks,” Building Environ., vol. 205, 2021, Art no. 108178. https://doi.org/10.1016/j.buildenv.2021.108178.
- [3]↑
B. Xia, X. Li, H. Shi, S. Chen, and J. Chen, “Style classification and prediction of residential buildings based on ML,” J. Asian Architecture Building Eng., vol. 19, no. 6, pp. 714–730, 2020. https://doi.org/10.1080/13467581.2020.1779728.
- [4]↑
Y. Zhou and H.-J. Park, “Sketch with artificial intelligence (AI) - a multimodal AI approach for conceptual design,” in PROJECTIONS - Proceedings of the 26th CAADRIA Conference, The Chinese University of Hong Kong, Hong Kong, 29 March – 1 April 2021, pp. 201–210. https://doi.org/10.52842/conf.caadria.2021.1.201.
- [5]↑
E. Cantemir and O. Kandemir, “Use of artificial neural networks in architecture: determining the architectural style of a building with a convolutional neural network,” Neural Comput. Appl., vol. 36, pp. 6195–6207, 2024. https://doi.org/10.1007/s00521-023-09395-y.
- [6]↑
M. Tsigkari, S. Tarabishy, and M. Kosicki, “Towards Artificial Intelligence in Architecture: how ML can change the way we approach design,” 2021. https://www.fosterandpartners.com/plus/towards-artificial-intelligence-in-architecture/. Accessed: May 19, 2023.
- [7]↑
K. Karabagli, M. Koc, P. Basu, and I. As, “A machine learning approach to translate graph representations into conceptual massing models,” in Proceedings of the XXV International Conference of the Ibero-American Society of Digital Graphics (SIGraDi), Online, 8 – 12 November 2021, pp. 191–202.
- [8]↑
A. Boim, J. Dortheimer, and A. Sprecher, “A machine-learning approach to urban design interventions, in: non-planned settlements,”, in POST-CARBON - Proceedings of the 27th CAADRIA Conference, Sydney, 9–15 April 2022, pp. 223–232. https://doi.org/10.52842/conf.caadria.2022.1.223.
- [9]↑
R. Deshpande, M. Nisztuk, C. Cheng, R. Subramanian, T. Chavan, C. Weijenberg, and S. V. Patel, “Synthetic machine learning for real-time architectural daylighting prediction,” in Proceedings of the 27th CAADRIA Conference, Sydney, 9–15 April 2022, pp. 313–322.
- [10]↑
CUMINCAD, Online open access index platform for publications in Computer Aided Architectural Design. https://papers.cumincad.org/.
- [11]↑
B. A. Zandavali and M. J. García, “Manuel automated brick pattern generator for robotic assembly using machine learning and images,” in Proceedings of the 37th eCAADe and 23rd SIGraDi Conference, University of Porto, Porto, Portugal, 11–13 September 2019, pp. 217–226.
- [12]↑
A. Andreou, O. Kontovourkis, S. Solomou, and A. Savvides, “Rethinking architectural design process using integrated parametric design and machine learning principles,” in Proceedings of the 41st Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe), Graz, 20–22 September 2023, pp. 461–470. https://doi.org/10.52842/conf.ecaade.2023.2.461.
- [13]↑
E. Bauscher, A. Dai, D. Elshani, and T. Wortmann, “Learning and generating spatial concepts of modernist architecture via graph ML,” in Proceedings of the 29th CAADRIA Conference, Singapore, vol. 1, 20–26 April 2024, pp. 159–168.
- [14]↑
C. Liu, J. Shen, Y. Ren, and H. Zheng, “Pipes of AI – machine learning assisted 3D modeling design,” in Proceedings of the 2020 DigitalFUTURES The 2nd International Conference on Computational Design and Robotic Fabrication (CDRF 2020), 2020. https://doi.org/10.1007/978-981-33-4400-6_2.
- [15]↑
J. Dong, S. Lin, and J. van Ameijde, “Predicting network integration based on satellite imagery around high-density public housing estates through machine learning,” in Proceedings of the XXVII International Conference of the Ibero-American Society of Digital Graphics (SIGraDi 2023), Punta del Este, Maldonado, Uruguay, 29 November – 1 December 2023, pp. 795–806. https://papers.cumincad.org/cgi-bin/works/Show?sigradi2023_387.
- [16]↑
G. Guida, “Multimodal architecture: applications of language in a machine learning aided design process,” in Proceedings of the 28th CAADRIA Conference, Ahmedabad, 18–24 March 2023, pp. 561–570. https://doi.org/10.52842/conf.caadria.2023.2.561.
- [17]↑
O. Z. Güzelci, “Machine learning in predicting section drawings - case of anatolian Seljuk Kümbets,” in Proceedings of the 40th Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2022), vol. 2, Ghent, 13–16 September 2022 pp. 169–176. https://doi.org/10.52842/conf.ecaade.2022.2.169.
- [18]↑
H. Zhang and Y. Huang, “Machine learning aided 2D-3D architectural form finding at high resolution,” in Proceedings of the 2020 DigitalFUTURES the 2nd International Conference on Computational Design and Robotic Fabrication (CDRF 2020), 2020. https://doi.org/10.1007/978-981-33-4400-6_15.
- [19]↑
S. Huang and H. Zheng, “Morphological regeneration of the industrial waterfront based on ML,” in Proceedings of the 27th CAADRIA Conference, Sydney, 9–15 April 2022, pp. 475–484. https://doi.org/10.52842/conf.caadria.2022.1.475.
- [20]↑
A. Sebestyen, J. Rock, and U. L. Hirschberg, “Towards Abductive Reasoning-Based Computational Design Tools - using Machine Learning as a way to explore the combined design spaces of multiple parametric models,” in Proceedings of the 39th eCAADe Conference, University of Novi Sad, Novi Sad, Serbia, 8–10 September 2021, pp. 141–150. https://doi.org/10.52842/conf.ecaade.2021.1.141.
- [21]↑
J. Shen, C. Liu, y. Ren, and H. Zheng, “Machine learning assisted urban filling,” in Proceedings of the 25th CAADRIA Conference, Chulalongkorn University, Bangkok, Thailand, 5–6 August 2020, pp. 679–688. https://doi.org/10.52842/conf.caadria.2020.2.679.
- [22]↑
H. M. Yau, T. Dounas, W. Jabi, and D. Lombardi, “Timber joints analysis and design using Shape and Graph Grammar based Machine Learning approach,” in Proceedings of the 41st Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe), Graz, 20–22 September 2023, pp. 569–578. https://doi.org/10.52842/conf.ecaade.2023.1.569.
- [23]↑
Y. Liu, C. Fang, Z. Yang, X. Wang, Z. Zhou, Q. Deng, and L. Liang, “Exploration on machine learning layout generation of Chinese private garden in Southern Yangtze,” in Proceedings of the 2021 DigitalFUTURES the 3rd International Conference on Computational Design and Robotic Fabrication (CDRF), 2021. https://doi.org/10.1007/978-981-16-5983-6_4.
- [24]↑
S. Zhang, Y. Li, S. Zhang, X. He, and R. Tian, “Text-to-Garden: generating traditional Chinese garden design from text-descriptions at scale with multimodal machine learning,” in Proceedings of the 28th CAADRIA Conference, Ahmedabad, 18–24 March 2023, pp. 79–88. https://doi.org/10.52842/conf.caadria.2023.1.079.
- [25]↑
Z. Dong and J. Lin, “Nolli map: interpretation of urban morphology based on machine learning,” in Proceedings of the 2022 DigitalFUTURES The 4st International Conference on Computational Design and Robotic Fabrication, (CDRF), 2022. https://doi.org/10.1007/978-981-19-8637-6_24.
- [26]↑
M. Albassel and M. M. Waly, “Applying machine learning to enhance the implementation of Egyptian fire and life safety code in mega projects architecture in the age of disruptive technologies,” in 9th ASCAAD Conference Proceedings, Cairo, Egypt [Virtual Conference] 2–4 March 2021, pp. 7–22. https://papers.cumincad.org/cgi-bin/works/Show?ascaad2021_021.
- [27]↑
S. Cao and H. Zheng, “A POI-based machine learning method in predicting health,” in Proceedings of the 41st Annual Conference of the Association of Computer Aided Design in Architecture (ACADIA), 3–6 November 2021, pp. 160–169. https://doi.org/10.52842/conf.acadia.2021.160.
- [28]↑
Z. Jiaxin, L. Yunqin, L. Haiqing, and W. Xueqiang, “Sensitivity analysis of thermal performance of granary building based on machine learning,” in Proceedings of the 24th CAADRIA Conference, Victoria University of Wellington, Wellington, New Zealand, 15–18 April 2019, pp. 665–674. https://doi.org/10.52842/conf.caadria.2019.1.665.
- [29]↑
K. Kharbanda, I. Papadopoulou, P. Pouliou, K. Daw, A. Belwadi, and H. Loganathan, “Learn Carbon - a tool for machine learning prediction of global warming potential from abstract designs,” in Proceedings of the 40th Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe), Ghent, 13–16 September 2022, pp. 601–610. https://doi.org/10.52842/conf.ecaade.2022.2.601.
- [30]↑
J. B. Kim, S. Kim, and J. Aman, “An urban building energy simulation method integrating parametric BIM and machine learning,” in Proceedings of the 28th CAADRIA Conference, Ahmedabad, 18–24 March 2023, pp. 665–674. https://doi.org/10.52842/conf.caadria.2023.1.665.
- [31]↑
F. M. La Russa and C. Santagati, “From the Cognitive to the Sentient Building - machine Learning for the preservation of museum collections in historical architecture,” in Proceedings of the 38th eCAADe Conference, TU Berlin, Berlin, Germany, 16–18 September 2020, pp. 507–516. https://doi.org/10.52842/conf.ecaade.2020.2.507.
- [32]↑
Y. Li, C. Huang, and J. Yao, “Optimising the control strategies for performance-driven dynamic building facades using machine learning,” in Proceedings of the 28th CAADRIA Conference, Ahmedabad, 18–24 March 2023, pp. 199–208. https://doi.org/10.52842/conf.caadria.2023.1.199.
- [33]↑
Y. Li, C. Huang, G. Zhang, and J. Yao, “Machine learning modeling and genetic optimization of adaptive building facade towards the light environment,” in Proceedings of the 27th CAADRIA Conference, Sydney, 9–15 April 2022, pp. 141–150. https://doi.org/10.52842/conf.caadria.2022.1.141.
- [34]↑
C. L. Lorenz, S. B. De Souza, and S. Packianather, “Machine learning in design exploration: an investigation of the sensitivities of ANN-based daylight predictions,” in 18th International Conference, CAAD Futures, Daejeon, Korea, 2019, pp. 75–87. https://papers.cumincad.org/cgi-bin/works/Show?cf2019_010.
- [35]↑
P. Nicholas, Y. Chen, N. Borpujari, N. Bartov, and A. Refsgaard, “A chained machine learning approach to motivate retro-cladding of residential buildings, Stojakovic,” in Proceedings of the 39th eCAADe Conference, University of Novi Sad, Novi Sad, Serbia, 8–10 September 2021, pp. 55–64. https://doi.org/10.52842/conf.ecaade.2021.1.055.
- [36]↑
A. Nogueira and L. Romao, “room_ID: an architectonic image classifier tool correlating machine learning and the domestic space,” in Proceedings of the XXVII International Conference of the Ibero-American Society of Digital Graphics (SIGraDi 2023), Punta del Este, Maldonado, Uruguay, 29 November - 1 December 2023, pp. 1833–1843. https://papers.cumincad.org/cgibin/works/Show?sigradi2023_417.
- [37]↑
N. Raanan, H. Yoffe, and J. Grobman, “A ML evaluation method for sustainability evaluation: the case of Neighbourhoods’ design,” in Proceedings of the 27th CAADRIA Conference, Sydney, 9–15 April 2022, pp. 283–291. https://doi.org/10.52842/conf.caadria.2022.1.283.
- [38]↑
A. Sepúlveda, N. Eslamirad, and F. De Luca, “Machine learning approach versus prediction formulas to design healthy dwellings in a cold climate,” in Proceedings of the 41st Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2023), Graz, 20–22 September 2023, pp. 359–368. https://doi.org/10.52842/conf.ecaade.2023.2.359.
- [39]↑
A. Sepúlveda, N. Eslamirad, S. Salehi, S. Shahabaldin, M. Thalfeldt, and F. De Luca, “Machine learning-based optimization design workflow based on obstruction angles for building facades,” in Proceedings of the 9th eCAADe Regional International Symposium, TalTech, 15–16 June 2023, pp. 15–24. https://papers.cumincad.org/cgibin/works/Show?ecaaderis2023_11.
- [40]↑
Y. Shi, J. Chen, G. Hu, and C. Wang, “Prediction and optimisation of the typical airport terminal corridor façade shading using integrated machine learning and evolutionary algorithms,” in Proceedings of the 28th CAADRIA Conference, Ahmedabad, 18–24 March 2023, pp. 59–68. https://doi.org/10.52842/conf.caadria.2023.1.059.
- [41]↑
H. Tu, “Eyes on the street: assessing window-to-wall ratios in Google street views using machine learning,” in Proceedings of the XXVI Conference of the Iberoamerican Society of Digital Graphics (SIGraDi 2022), Universidad Peruana de Ciencias Aplicadas, Lima, 7–11 November 2022, pp. 175–186. https://papers.cumincad.org/cgibin/works/Show?sigradi2022_35.
- [42]↑
Q. M. Xu, G. d. C. Lopez, J. Luis, and H. W. Samuelson, “Towards a decision framework integrating physics-based simulation and machine learning in conceptual design,” in Proceedings of the 28th CAADRIA Conference, Ahmedabad, 18–24 March 2023, pp. 371–380. https://doi.org/10.52842/conf.caadria.2023.2.371.
- [43]↑
H. Ahmadpanah, A. Haidar, and S. M. Latifi, “BIM and machine learning (ML) integration in design coordination: using ML to automate object classification for clash detection,” in Proceedings of the 41st Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2023), Graz, 20–22 September 2023, pp. 619–628. https://doi.org/10.52842/conf.ecaade.2023.2.619.
- [44]↑
A. Alymani, W. Jabi, and P. Corcoran, “Machine learning methods for clustering architectural precedents - classifying the relationship between building and ground,” in Proceedings of the 38th eCAADe Conference, TU Berlin, Berlin, Germany, 16–18 September 2020 2020, pp. 643–652. https://doi.org/10.52842/conf.ecaade.2020.1.643Aly.
- [45]↑
G. Austern, R. Yosifof, and D. Fisher-Gewirtzman, “A dataset for training machine learning models to analyze urban visual spatial experience,” in Proceedings of the 41st Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2023), Graz, 20–22 September 2023, pp. 781–790. https://doi.org/10.52842/conf.ecaade.2023.2.781.
- [46]↑
L. Brown, M. Yip, N. Gardner, M. H. Haeusler, N. Khean, Y. Zavoleas, and C. Ramos, “Drawing recognition - integrating ML systems into architectural design workflows,” in Proceedings of the 38th eCAADe Conference, TU Berlin, Berlin, Germany, 16–18 September 2020, pp. 289–298. https://doi.org/10.52842/conf.ecaade.2020.2.289.
- [47]↑
D. Liu and K. Wang, “Spatial analysis of villages in Jilin Province based on space syntax and machine learning,” in Proceedings of the 2022 DigitalFUTURES the 4st International Conference on Computational Design and Robotic Fabrication (CDRF), 2022. https://doi.org/10.1007/978-981-19-8637-6_1.
- [48]↑
C. Ferrando, N. Dalmasso, J. Mai, and D. C. Llach, “Architectural distant reading using machine learning to identify typological traits across multiple buildings,” in 18th International Conference, CAAD Futures, 2019, pp. 114–127. https://papers.cumincad.org/cgibin/works/Show?cf2019_014.
- [49]↑
M. Jiang and C. Cai, “Communication with Detroit: machine learning in open source community housing design,” in Proceedings of the 28th CAADRIA Conference, Ahmedabad, 18–24 March 2023 2023, pp. 49–58. https://doi.org/10.52842/conf.caadria.2023.1.049.
- [50]↑
A. Meeran and S. Conrad Joyce, “Machine learning for comparative urban planning at scale: an aviation case study,” in Proceedings of the 40th Annual Conference of the Association of Computer Aided Design in Architecture (ACADIA), 2020, pp. 178–187. https://doi.org/10.52842/conf.acadia.2020.1.178.
- [51]↑
D. Raghu, A. Markopoulou, M. Marengo, I. Neri, A. Chronis, and C. De Wolf, “Enabling component reuse from existing buildings through machine learning, using Google street view to enhance building databases,” in Proceedings of the 27th CAADRIA Conference, Sydney, 9–15 April 2022, pp. 577–586. https://doi.org/10.52842/conf.caadria.2022.2.577.
- [52]↑
W. Ran, L. Yin, and J. Yu, “Machine learning-driven comparative study: morphological taxonomy in screen-based building clusters,” in Proceedings of the 43rd Annual Conference for the Association for Computer Aided Design in Architecture (ACADIA), 2023, pp. 596–605. https://papers.cumincad.org/cgi-bin/works/Show?acadia23_v2_596.
- [53]↑
J. Rhee, D. C. Llach, and R. Krishnamurti, “Context-rich urban analysis using machine learning - a case study in Pittsburgh,” in Proceedings of the 37th eCAADe and 23rd SIGraDi Conference, University of Porto, Porto, Portugal, 11–13 September 2019, pp. 343–352. https://doi.org/10.52842/conf.ecaade.2019.3.343.
- [54]↑
X. Wang, P. Tang, and C. Cai, “Traditional Chinese village morphological feature extraction and cluster analysis based on multi-source data and machine learning,” in Proceedings of the 28th CAADRIA Conference, Ahmedabad, 18–24 March 2023, pp. 179–188. https://doi.org/10.52842/conf.caadria.2023.1.179.
- [55]↑
Y. Wang and X. Li, “The AI’s Cognition of the Cultural Characteristics Underlying Chinese Courtyard Dwelling: a visual perception-based approach for Siheyuan plan classification using machine learning and convolutional neural networks,” in Proceedings of the 42nd Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2024), Nicosia, vol. 1, 11-13 September 2024, pp. 643–652. https://doi.org/10.52842/conf.ecaade.2024.1.643.
- [56]↑
X. Xia and Z. Tong, “A machine learning-based method for predicting urban land use,” in Proceedings of the 25th CAADRIA Conference, Chulalongkorn University, Bangkok, Thailand, 5–6 August 2020, pp. 21–30. https://doi.org/10.52842/conf.caadria.2020.2.021.
- [57]↑
X. Zhao, T.-H. Wang, and C. Peng, “Automatic room type classification using machine learning for two-dimensional residential building plans”, in Proceedings of the 40th Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2022), Ghent, 13–16 September 2022, pp. 593–600. https://doi.org/10.52842/conf.ecaade.2022.2.593.
- [58]↑
H. Zheng and Y. Ren, “Machine learning neural networks construction and analysis in vectorized design drawings,” in Proceedings of the 25th CAADRIA Conference, Chulalongkorn University, Bangkok, Thailand, 5–6 August 2020, p. 707. https://doi.org/10.52842/conf.caadria.2020.2.707.
- [59]↑
M. Tamke, P. Nicholas, and M. Zwierzycki, “Machine learning for architectural design: practices and infrastructure”. Int. J. Architect. Comput., vol. 16 no. 2, pp. 123–143, 2018. https://doi.org/10.1177/1478077118778580.
- [60]↑
S. K. Baduge, S. Thilakarathna, J. S. Perera, M. Arashpour, P. Sharafi, B. Teodosio, A. Shringi, and P. Mendis, “Artificial intelligence and smart vision for building and construction 4.0: machine and deep learning methods and applications,” Automation Construct., vol. 141, 2022, Art no. 104440. https://doi.org/10.1016/j.autcon.2022.104440.
- [61]↑
M. del Campo, “Architecture, language and AI - language, attentional generative adversarial networks (AttnGAN) and architecture design,” in Proceedings of the 26th CAADRIA Conference, The Chinese University of Hong Kong and Online, Hong Kong, 29 March - 1 April 2021, pp. 211–220. https://doi.org/10.52842/conf.caadria.2021.1.211.
- [62]↑
G. Özel, “Interdisciplinary AI: a ML system for streamlining external aesthetic and cultural influences in architecture,” in Architectural Intelligence, P. F. Yuan, M. Xie, N. Leach, J. Yao, and X. Wang, Eds., Singapore: Springer, 2020. https://doi.org/10.1007/978-981-15-6568-7_7.
- [63]↑
A. Andreou, O. Kontovourkis, S. Solomou, and A. Savvides, “Rethinking architectural design process using integrated parametric design and machine learning principles,” in Proceedings of the 41st Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2023), Graz, 20–22 September 2023, pp. 461–470. https://doi.org/10.52842/conf.ecaade.2023.2.461.
- [64]↑
I. Karadag, “Machine learning for conservation of architectural heritage,” Open House Int., vol. 48, no. 1, pp. 23–37, 2023. https://doi.org/10.1108/OHI-05-2022-0124.
- [65]↑
Y. Miky, Y. Alshawabkeh, and A. Baik, “Using deep learning for enrichment of heritage BIM: Al Radwan house in historic Jeddah as a case study,” Herit Sci., vol. 12, p. 255, 2024. https://doi.org/10.1186/s40494-024-01382-3.
- [66]↑
N. Karimi, M. Mishra, and P. B. Lourenço, “Deep learning-based automated tile defect detection system for Portuguese cultural heritage buildings,” J. Cult. Heritage, vol. 68, pp. 86–98, 2024. https://doi.org/10.1016/j.culher.2024.05.009.
- [67]↑
L. M. Obeidat, M. H. Wedyan, T. S. Al-Radaideh, S. N. Ma’bdeh, and R. Shannik, “A combination of ML algorithms and parametric design tools to predict the energy consumption for residential buildings,” Int. Rev. Civil Eng. (IRECE), vol. 15, no. 2, 2024. https://doi.org/10.15866/irece.v15i2.23601.
- [68]↑
A. Alymani, W. Jabi, and P. Corcoran, “Graph, “ML classification using architectural 3D topological models”,” SIMULATION, vol. 99, no. 11, pp. 1117–1131, 2023. https://doi.org/10.1177/00375497221105894.
- [69]↑
K. Tochaiwat and P. Seniwong, “House type specification for housing development project using ML techniques: a study from Bangkok metropolitan region, Thailand,” Nakhara: J. Environ. Des. Plann., vol. 23, no. 1, 2024, Article 403.
- [70]↑
F. C. Kim and J. Huang, “Deep architectural archiving (DAA), towards a machine understanding of architectural form,” in Proceedings of the 27th CAADRIA Conference, Sydney, 9–15 April 2022, pp. 727–736. https://doi.org/10.52842/conf.caadria.2022.1.727.
- [71]↑
E. Millan, M.-V. Belmonte, F.-J. Boned, J. Gavilanes, et al., “Using ML techniques for architectural design tracking: an experimental study of the design of a shelter,” J. Building Eng., vol. 51, no. 104223, 2022. https://doi.org/10.1016/j.jobe.2022.104223.
