AI-enhanced conversational design processes: A systematic review

Abstract

AI technologies have been widely explored in the architectural design process over the last decade. This paper addresses the limitations of the existing reviews on AI applications in the architectural design process, including a lack of focus on the design process, limited coverage of AI technologies, and under-exploration of human-AI interaction environments. The paper systematically reviews and comparatively analyses 63 articles filtered through 1138 publications from 8 databases. The findings include comparative analysis charts and tables of the different AI-enhanced design processes, the human-AI interaction environment, as well as their evaluation. They highlight the expectation that AI will function as a conversational design partner. Lastly, the paper presents a novel framework for the AI-enhanced conversational architectural design process.

Keywords

conversational design process architectural design process artificial intelligence design evaluation design framework

Introduction

Artificial Intelligence (AI) tools are widely explored in the architectural design process, especially in the conceptual design process.¹ There are existing systematic reviews of AI applications in the architecture field. Three observations are observed, including (1) a lack of design process focus, (2) limited or outdated coverage of AI technologies and (3) little focus on human-AI interaction.

While systematic reviews exist on AI applications across Architecture, Engineering, and Construction (AEC), a notable gap remains in examining its role within the architectural design process. Momade et al.² conducted a systematic review of AI applications in the AEC industry, offering researchers and practitioners a comprehensive understanding of AI’s role. However, their studies prioritised cost optimisation through predictive AI during modelling, paying limited attention to design-related applications. Furthermore, their broad AEC scope does not align with this study, which centres on architectural design processes. Similarly, Saka et al.³ emphasised conversational AI and natural language processing (NLP) within the AEC industry. As their review targeted industry-wide productivity and efficiency improvements rather than architectural design, it also falls outside the scope of this review. Rane et al.⁴ examined leading-edge technologies for architectural design, covering innovations such as immersive technologies (VR, AR), digital twins, drones, and digital fabrication. While that study emphasised collaboration, it focused on architect-stakeholder collaboration via AI-enabled management systems rather than AI applications directly within the design process.

Furthermore, existing systematic reviews address AI in the architectural design process, yet they focus on a restricted set of tools and often outdated technologies. Castro Pena et al.¹ examined the role of AI in the conceptual design phase of architecture, concentrating on optimisation algorithms with some mention of artificial neural networks (ANNs). However, the study is outdated, as it does not reflect the significant advancements in generative AI, which have become prominent since 2021 and offer a more substantial influence on the design process. Also, Enjellina et al.⁵ review of text-to-image AI generators and their application in the architectural field, examining their influences, challenges and future potential. The study maintained a narrow scope by exclusively addressing text-to-image AI, omitting other AIs being discussed in other reviews, such as large language models and regression models.

At last, despite there being systematic reviews on how the designers interact with the AI tools, little focus is on how designers interact with the computational or AI tools. For instance, Li et al.⁶ focus on generative AI in the architectural design process, focusing on AI models that create visual outputs such as 2D images, 3D models, and videos. The review prioritised analysing the technical aspects of the AI models, such as their mathematical mechanisms and different version comparisons, rather than exploring how designers are practically utilising them in their workflows. Vissers-Similon et al.⁷ reviewed and classified AI techniques for early architectural design stages, focusing on the technology and its applications, including recent AI developments, but did not delve into how designers interact with these computational tools. Table 1 summarises the reviews of AI applications in the architecture, engineering and construction fields.

Table 1.

Reviews of AI in the architecture field.

	Domain	Focus	Limitation
Momade et al., 2021²	AEC	Cost optimisation by applying predictive AI during the modelling process	Not focused on the design process
Saka et al., 2023³	AEC	Application of natural language processing (NLP) in productivity efficiency improvement
Rane et al., 2023⁴	Architecture	A wide range of technologies (immersive technologies, digital twin, drones, digital fabrication), focusing on architect-stakeholder collaborations through an AI-enabled management system
Castro Pena et al., 2021¹		Optimisation algorithms and artificial neural network (ANN) application in the conceptual design stage	Outdated in that it does not include the latest generative AIs due to the year of research and publication
Enjellina et al., 2023⁵		Text-to-image-based AI image generators for the architecture discipline	Not including AI technology, except for AI image generators
Li et al., 2024⁶		AI 2D, 3D and video generators in the architectural design process, with an exploration of AI’s (algorithmic) mechanism and “versions” of different AI models	Not including how designers interact with the computational/AI tools
Vissers-Similon et al., 2024⁷		Overall view of technological applications in the architectural design process, including the latest AI development

Therefore, the overarching research question in this study is: How can AI be used in the architectural design process? Consequently, this systematic review focuses on three aspects: (1) AI application in the early-stage architectural design processes, (2) human-AI interaction environments (code-based environment, software, web application, etc.), and (3) evaluation of the designer-AI interaction. It is noted that for the first aspect, AI refers to a broad spectrum from regression-based, convolutional neural network (CNN)/generative adversarial network (GAN)-based to Recurrent Neural Network (RNN)/transformer-based systems, not limited to generative AIs such as image generators or chatbots.

For the third aspect, the term “evaluation” is considered in two dimensions. Firstly, it refers to what is being evaluated, which can include (a) AI performance (e.g. AI prediction accuracy or AI operation efficiency), (b) the quality and diversity of design outputs generated with AI, and (c) the human-AI design process itself, including designers’ experience and observed design behaviours. Second, it refers to how evaluation is conducted, ranging from efficiency-focused metrics, through demonstration-based case studies, to workshop experimentation (case studies) that combine demonstration with quantitative measures (e.g. surveys, counts of usable AI outputs or success rates, etc.) and qualitative methods (e.g. observations and interviews, etc.). This paper takes the approach of answering the following sub-questions:

(1) How can we utilise AI technology in the conversational architectural design process?

(2) Through which interaction environments do designers interact with AI during the design process?

(3) How have existing studies evaluated the applications of AI in architectural design?

Findings include a comparative analysis chart evaluating AI technologies and applications in the architectural design process and the human-AI interaction environments. Then, key principles for the AI-enhanced architectural design process are summarised for future research and design development.

Methodology

The systematic review method is composed of four phases: (1) Searching articles, (2) Filtering articles, (3) Comparative analysis and categorisation of articles and (4) Evaluation in charts and tables (Figure 1).

Figure 1.

Flowchart of the systematic review method.

Phase 1: Searching articles through the database

Table 2 illustrates the search procedure with sources of databases and respective search strings. The search strings are comprised of three parts: (1) artificial intelligence, (2) architectural design, and (3) design process. To ensure the inclusion of relevant publications, the keywords are further expanded with (1) AI, machine learning, deep learning, (2) architecture design, building design and (3) design exploration. The search query was made to eight databases, including common databases such as Scopus (Elsevier’s abstract and citation database), Semantic, Sage Journals and Web of Science Core Collection. In addition, more engineering-based databases are included, such as ScienceDirect, Scholar, IEEE Xplore (Institute of Electrical and Electronics Engineers), and ACM (Association of Computing Machinery) Digital Library. Finally, CumInCAD (Cumulative Index about publications in Computer Aided Architectural Design) is included since it covers computational architecture research conferences, including ACADIA, CAADRIA, eCAADe, SIGraDi, ASCAAD and CAAD futures. The search results between 2014 and 2024 are collected.

Table 2.

Search procedure performed in databases.

Source	Search string
Scopus	(“AI” OR “artificial intelligence” OR “machine learning” OR “deep learning”) AND (“architectural design” OR “architecture design” OR “building design”) AND (“design process” OR “design exploration”) in Article title, Abstract, Keywords Filter: Year: 2014-2024 Filter: Engineering, Computer Science, Social Sciences, Arts and Humanities, Environmental Science, Multidisciplinary
ScienceDirect, Semantic Scholar, IEEE Xplore	(“AI” OR “artificial intelligence” OR “machine learning” OR “deep learning”) AND (“architectural design” OR “architecture design” OR “building design”) AND (“design process” OR “design exploration”) Filter: Year: 2014-2024 Filter: Only content I have full access to
Sage Journals	(“AI” OR “artificial intelligence” OR “machine learning” OR “deep learning”) AND (“architectural design” OR “architecture design” OR “building design”) AND (“design process” OR “design exploration”) Filter: Year: 2014-2024 Filter: Abstract Excluded
Web of Science Core Collection	(“AI” OR “artificial intelligence” OR “machine learning” OR “deep learning”) AND (“architectural design” OR “architecture design” OR “building design”) AND (“design process” OR “design exploration”) Filter: Year: 2014-2024 Filter: Architecture
ACM Digital Library	[[All: “ai”] OR [All: “artificial intelligence”] OR [All: “machine learning”] OR [All: “deep learning”]] AND [[All: “architectural design”] OR [All: “architecture design”] OR [All: “building design”]] AND [[All: “design process”] OR [All: “design exploration”]] AND [E-Publication Date: (01/01/2014 TO 12/31/2024)]
CumInCAD	Due to the limitations of the CumInCAD search engine, a reduced search string has been used: {summary} = ∼ m/AI/i and {summary} = ∼ m/architectural design process/i Filter: Year: 2014-2024

The search results included 1138 academic publications, including 164 publications in Scopus, 22 publications in ScienceDirect, 58 articles in SAGE Journals, 10 publications in Semantic Scholar, 60 publications in IEEE Xplore, 28 publications in Web of Science Core Collection, 725 publications in ACM Digital Library and 71 conference papers in CumInCAD.

Phase 2: Filtering articles

The literature selection process is summarised in Figure 2. A total of 1138 studies were searched in the initial search. After removing 39 duplicates, 1099 studies remained. After excluding non-article records (abstracts, books, and proceedings), 756 studies remained. After filtering based on the inclusion and exclusion criteria displayed in Table 3 based on the abstract reading, 70 studies were included. Then, after reading the full papers, 60 studies were chosen. During the review process, snowball papers are included, making a total of six studies in the final synthesis.

Figure 2.

Literature selection process.

Table 3.

Inclusion and exclusion criteria.

Inclusion criteria	Exclusion criteria
Architecture Discipline	Not in the architecture discipline
Focused on the design stage	Not focused on the design stage (e.g. construction, manufacturing stage)
AI tools are used	Did not involve any AI tools
Architecture or urban design	Design in other disciplines (e.g. industrials)
Peer-reviewed	No peer-reviewed
Empirical studies	Literature reviews, meta-analysis, etc.
Written in English	Written in languages other than English

Phase 3: Comparative analysis and categorisation of articles

In Phase Three, comparative analysis and categorisations are conducted according to the three sub-research questions. First, articles are categorised and analysed based on three big categories of AIs and their application in the architectural design process, responding to the first sub-research questions. Then, for each category, analysis based on interaction environments of how humans and AI interact and how the data of the AI-integrated architectural design process is evaluated are comparatively analysed, responding to research questions 2 and 3, respectively.

Phase 4: Evaluation in charts and tables

In the final phase, the articles are evaluated in charts and tables. Based on the findings of phase four, an AI-enhanced architectural design framework for early-stage exploration is proposed.

Current status of the existing research

Distribution of articles according to the publication source

Table 4 shows the article distribution based on publication sources. Over 22% of the AI-integrated architectural design process-related publications are mainly from CAADRIA (Computer-Aided Architectural Design Research in Asia) and eCAADe (International Conference on Education and Research in Computer-Aided Architectural Design in Europe), followed by journals which are focusing on computer-aided design and engineering, including the International Journal of Architectural Computing and Journal of Building Engineering, which occupy nearly 15% of the overall publications. It displays this topic’s awareness across multi-disciplinary research beyond architectural design, but extended to computing, building engineering and education.

Table 4.

Number of articles published on respective data sources.

Journal/Book/Conference proceedings	# of articles
International Conference on Education and Research in Computer-Aided Architectural Design in Europe (eCAADe)	8
Computer-Aided Architectural Design Research in Asia (CAADRIA)	7
International Journal of Architectural Computing	5
Journal of Building Engineering	4
Association for Computer Aided Design in Architecture (ACADIA)	3
Buildings	2
Arab Society for Computer-Aided Architectural Design (ASCAAD)	2
Automation in Construction	2
Proceedings of the Symposium on Simulation for Architecture & Urban Design	2
Simulation Series	2
Iberoamerican Society of Digital Graphics (SiGraDi)	1
Frontiers of Architectural Research	1
IEEE/CVF International Conference on Computer Vision (ICCV)	1
Advances in Architectural Geometry	1
Advances in Intelligent Systems and Computing	1
Building and Environment	1
Computer-Aided Architectural Design Futures (CAAD Futures)	1
Computational Design and Robotic Fabrication (CDRF)	1
CEUR Workshop Proceedings	1
Civil Engineering and Architecture	1
Conference on Creativity and Cognition	1
Design Computation Input/Output	1
Design Journal	1
Digital Government: Research and Practice	1
Engineering Structures	1
Frontiers in Artificial Intelligence and Applications	1
IEEE Transactions on Visualisation and Computer Graphics	1
International Conference on Computing in Civil and Building Engineering	1
International Conference on Consumer Electronics and Computer Engineering (ICCECE)	1
International Congress on Human-Computer Interaction, Optimisation and Robotic Applications (HORA)	1
Journal of Computational Design and Engineering	1
Proceedings of Intelligent Computing in Engineering	1
Proceedings of the Genetic and Evolutionary Computation Conference	1
Science and Information Conference	1
Towards Radical Regeneration: Design Modelling Symposium Berlin 2022	1

CumInCAD sibling associations (CAADRIA, eCAADe, ACADIA, ASCAAD, SiGraDi, CAAD Futures, and CDRF) contribute more than 35% of the publications. This indicates its importance and relevance in the field of computer-aided architectural design. The articles appear in a total of 36 different journals, books, or conference proceedings, reflecting the topic’s broad coverage across fields like architecture, design, human-computer interaction, engineering, and information science, displaying the potential for multi-disciplinary implications.

Distribution of articles according to publication year

All selected papers are published between 2014 and 2024 Figure 3. There is an increasing trend in the number of articles published over the years, especially after 2021, when generative AI tools have been advancing,¹ with an exception in 2022. The rising trend shows growing research interest and research relevance in this field.

Figure 3.

Chronological distribution of articles.

Analysis and categorisation

Since it is observed that the sudden trend is closely related to the development of AIs, the first emphasis is placed on the types of AI technologies and their respective applications. There are two categories: (1) regression-based, which refers to numeric prediction AIs, and (2) generative AIs. Generative AI can be subdivided into two main categories: convolutional neural network (CNN) or generative adversarial network (GAN)-based, which includes the early development of image recognition, image generation, such as Deepdream and Style Transfer; and transformer-based, which includes chatbots and more common image generations, such as Stable Diffusion and Midjourney.⁸

This regression/convolutional neural network (CNN)/Recurrent Neural Network (RNN) categorisation is referenced to the syllabus established by computer scientist Andrew Ng, who set up the first machine learning course for Stanford University before 2021. The course was later published on the MOOC platform DeeplearningAI as free courses, titled “Machine Learning Specialisation” and “Deep Learning Specialisation”, which over 1.4 million people enrolled in at the time of Jan 2025.^9,10

Regression-based AI

The first category consists of AIs trained for numerical prediction tasks, including methods such as artificial neural networks (ANNs), random forests, swarming algorithms. Their applications can be grouped into three main areas: (1) enhancing performative design efficiency, (2) generating 3D models and (3) supporting design analysis and evaluation.

The most common application was to use AI to train surrogate models for efficient or real-time simulation to enhance performative design efficiency. For example, in the massing model design stage,¹¹ Sebestyen & Tyc, Jakub¹² and Lin & Tsay¹³ applied ANN as a surrogate model for more efficient daylight simulation for massing models for applying in the early-stage architectural design process. Hasan & Horvat¹⁴ trained a random forest model to predict Solar Envelope’s Floor Space Index (FSI), and Salamanca et al.¹⁵ trained an autoencoder as a surrogate model to quickly estimate the performance attributes for the generated massing models. Such a surrogate model approach was also applied in design aspects other than the massing model. For instance, Rahmani Asl¹⁶ used ANN and Boosted Decision Tree (BDT) for rapid energy and daylight performance prediction for façade design. It was also explored in more site-specific, contextual design approaches. Wang et al.¹⁷ applied ANN to suggest the pedestrian entrance location of a building site, and Wu & Liu¹⁸ used ANN to analyse which parts of the site were good for public/private use by considering the distance from specific environmental features, specific zones and limited conditions. It showed that despite ANN being the main algorithm explored in massing model generations, other algorithms were used for more specific design feature predictions. In response to this phenomenon, in 2024, Ayman et al.¹⁹ tested 13 different regression and classification models to compare the performances between different regression-based models. Its findings showed that Random Forest and XGBoost performed the best. Such efficient applications of the prediction model described above allow simulation to shift from a post-design evaluation tool to a real-time design tool.

The second application was to use AI as a 3D model generator. Gerber et al.²⁰ developed a multi-agent system (MAS) framework for integrated and performance-driven architectural design. It is composed of (1) two sets of MAS modular-based and particle-based, and (2) an AI-enhanced voting system. For the former application, a set of MAS is used to design light diffusion panel modules, while another set of agents is used to design façade porosity. For the latter application, an AI voting system is proposed to enhance the communication between two MAS. Similarly, D. Wang & Snooks, Roland²¹ applied reinforcement learning with MAS, where agents serve as the design geometry (mesh-graph) themselves while evolving with design criteria (spatial coverage, topological performance, site response and hierarchical reward) as the reward for reinforcement learning.

Finally, AI was also used as a design analyser, which was explored to analyse design enquiry, design forms and design process, respectively. For example, Başarır & Erol, Kutluhan²² used Word2Vec to process and translate textual descriptions of briefs as a design enquiry analyser. Algeciras-Rodríguez²³ and Harding²⁴ trained Self-Organising Map (SOM) models for analysing architectural features of the design models as an abstraction of visual information; as such, the information could be used for regeneration as 3D inspiration for future design. Similarly, Kazemi et al.²⁵ applied random forest for architectural and structural features detection and a decision tree for exploring the correlations between architectural and structural parameters, and Kim et al.²⁶ used a clustering algorithm and Struct2Vector to analyse design forms and typology for early stage design. Finally, Millán et al.²⁷ applied WEKA software (based on random forest and Bayesian classifier) and a clustering algorithm for the design process to help identify common problem-solving (function and spatial) design patterns. Finally, Schaumann et al.²⁸ applied a swarming algorithm to simulate human user behaviour, including individual and group activities.

Table 5 summarises the results above. Half of the publications focused on exploring AI as a surrogate model to turn the role of design simulation from an evaluation to design. One-third of the research projects explored how AI can be used to analyse design enquiry, forms and processes. At the same time, the rest studied AI as architectural geometries or simulated users by developing multi-agent system frameworks, which have not been popular since 2015.

Table 5.

Summary of ANN/regression-based AI applications.

Category	Publication	AI technologies	Applications
AI as a surrogate model	Rahmani Asl, 2017¹⁶	ANN and BDT	Performative prediction
	Radziszewski & Waczyńska, 2018¹¹	ANN
	Sebestyen & Tyc, Jakub, 2020¹²	ANN
	Lin & Tsay, 2021¹³	ANN
	Salamanca et al., 2023¹⁵	Autoencoder
	Hasan & Horvat, 2024¹⁴	Random forest
	Z. Wang et al., 2018¹⁷	ANN	Contextual design prediction
	Wu & Liu, 2020¹⁸	ANN	Contextual design prediction
	Ayman et al., 2024¹⁹	13 regression and classification models	Surrogate model performance comparison
AI as a 3D model generator	Gerber et al., 2015²⁰	Swarming algorithm	Agents as architectural geometries and a voting system
AI as a 3D model generator	D. Wang & Snooks, Roland, 2021²¹	Reinforcement learning	Agents as architectural geometries
Design analyser	Başarır & Erol, Kutluhan, 2021²²	Word2Vec	Design enquiry analyser
	Algeciras-Rodríguez, 2016²³	SOM	Design form analyser
	Harding, 2016²⁴	SOM
	Kim et al., 2023²⁶	Random forest
	Kazemi et al., 2024²⁵	Clustering algorithm and Struct2Vector
	Millán et al., 2022²⁷	Random forest, Bayesian classifier and clustering algorithm	Design process analyser
	Schaumann et al., 2015²⁸	Swarming algorithm (MAS)	User behaviour simulations

Human-AI interaction environments

The review of the studies reveals that humans (both researchers and designers) primarily interacted with regression-based AIs through coding environments such as Python, Anaconda and MATLAB, as summarised in Figure 4. Although six studies did not specify the platforms used, from the AI model training and deployment descriptions and graphs in the publications, they were most probably trained and made inferences in code-based environments. There are five research studies that, after the AI models were trained in a coding environment, the application demonstrations were either conducted or claimed to be performed in architectural software. For example, five researchers deployed the trained models in Rhino and Grasshopper,^{11,15,17,19,24} while Rahmani Asl¹⁶ deployed the trained model at Autodesk Dynamo Studio. Two of the MAS research projects were solely conducted within the software with user interfaces, took advantage of gaming development software such as Unity, and experiments were demonstrated using the immersive virtual environment²⁰ and 3D software such as Unity 3D and 3DS Max.²⁸

Figure 4.

Environments of human and regression-based AI interaction from 2015 to 2024.

In summary, it remains inevitable for researchers to interact with regression-based AI models within coding environments, which require coding knowledge. However, there was no lack of attempts to transition the trained models into practice environments in which designers in the architecture discipline could interact more intuitively and lower the learning barrier.

Evaluation methods

For seven surrogate model efficiency-focused research papers, the AI model evaluation was based on comparing the trained AI model predictions to ground truths,^14,18,25 simulation data¹² or similarities between the AI selection with students’ choices. Wang et al.,¹⁷ Radziszewski & Waczyńska¹¹ and Lin & Tsay¹³ compared the trained surrogate model prediction time to traditional simulation methods. Three researchers took a step forward after evaluating the AI model accuracy; they demonstrated the applications and displayed images during the design process simulation.^15,16,19 Seven of the 18 research papers focused on illustrating the AI application in the design process with images and verbal descriptions, emphasising the impact on the design process.^{20–24,26,28} Only one of the 18 researchers applied the AI model after the design process. Millán et al.²⁷ conducted a shelter design workshop with 52 designers; the design process images were collected and analysed using the trained AI model.

As a result, it is evident that although AI model training is crucial, most evaluation methods now prioritise demonstrating or observing how AI is used in the design process, rather than solely focusing on the efficiency or accuracy of the trained AI models.

Generative AIs (CNN/GAN-based)

The second category is a convolutional neural network (CNN) or generative adversarial network (GAN)-based, which is mostly dedicated to image processing, including image recognition and generation. The applications are categorised into 2D image generation, 3D model generation, design generation and design analysis by image recognition, as summarised in Table 6.

Table 6.

Summary of CNN/GAN-based AI applications.

Category	Publication	AI technologies	Applications
AI as a 2D inspiration image generator	Başarır & Erol, Kutluhan, 2021²²	CGAN	Generate inspirational images from texts
	del Campo, 2021³¹	AttnGAN, GraphCNN
	Horvath & Pouliou, 2024³²	VQGAN, CLIP
	Danchenko, 2021³⁰	CNN, StyleGAN	Generate inspiration images from images
	Qian et al., 2023²⁹	Style Transfer	Generate inspiration images from images
AI as a 3D inspiration model generator	Bank, 2022³⁸	StyleGAN2	Generate 2D images for 3D (manual) modelling
	Liu et al., 2021³⁹	Style Transfer	Generate 2D images for 3D (manual) modelling
	del Campo et al., 2021³⁶	VGG-16 (based on CNN)	Generate 3D models from AI-generated 2D images
	Yousif & Bolojan, 2024³⁷	StyleGAN, Pix2pix, CycleGAN	Generate 3D models from AI-generated 2D images
	del Campo et al., 2020⁴⁴	GCNN (3D CNN)	Generate 3D models from the trained dataset
	Ennemoser & Mayrhofer-Hufnagl, 2023⁴⁵	DCGAN
	Pouliou et al., 2023⁴⁶	CPCGAN (3D GAN)
	Zhuang et al., 2023⁴⁷	GAN
	Rasoulzadeh et al., 2024³⁴	Stroke2Surface	Generate 3D models from a 3D sketch (on tablet)
	Ampanavos & Malkawi, 2022³⁵	Variational Autoender (VAE) (involve CNN)	Generate 3D from performative criteria.
AI as a design drawing generator	Eisenstadt, 2019⁴⁰	GAN	Generate floorplan
	Y. Zhang et al., 2024³³	Pix2pix	Generate floorplan
	Liao et al., 2021⁴¹	StructGAN	Generation structural plan
AI as a design analyser	Clever et al., 2021⁴²	CNN	Analyse floorplan
AI as a design analyser	Sabah et al., 2024⁴³	CoreML	Detect architectural elements (object detection)

For 2D image generation, two researchers generated inspiration images from the image input. Qian et al.²⁹ generated sketches according to shape preferences using Style Transfer, which allows designers to provide a design image and a “style” image as a reference that the “style” would be mapped onto the original design image. Danchenko³⁰ used CNN to classify the clients’ intention by their selected precedent image from a dataset with about one thousand images prepared by the designer, then the designer used StyleGAN to generate three mood board images. In addition, three research processes used text descriptions as inputs to generate images. While their image generation was based on GAN, text processing had different approaches. del Campo³¹ applied another GAN (AttnGAN) model to process the input texts while Başarır & Erol and Kutluhan²² used Word2Vec, a regression-based AI, and Horvath & Pouliou³² and Zhang et al.³³ used Recurrent Neural Network (RNN) and transformer-based AIs, which are introduced in the next category. Such observations show researchers’ awareness of connections between natural language and AI-enhanced design processes and display the potential to leverage the three types of AI technologies according to needs rather than finding the “best AI model” with the highest efficiency or prediction accuracy.

For 3D model generations, almost half of the 10 research proposals developed methodologies that generated 3D models from trained datasets, similar to regression-based AI approaches. The remaining six researchers explored 3D model generation from sketch,³⁴ performative criteria,³⁵ 2D design images^36,37 or proposed manual computational processing in 3D modelling software such as Rhinoceros and Grasshopper from a series of AI-generated 2D images.^38,39 This shows a trend of merging AI generation into the common design workflow, rather than treating it as a standalone AI application.

Despite GAN and CNN being novel technologies in generating 2D images or 3D models, there were researchers leveraging the generative properties of the AI models to generate architectural information, such as floor plans^33,40 or structural plans,⁴¹ instead of pixelated inspiration images. Others utilised GAN’s and CNN’s image recognition capability for floorplan analysis⁴² and architectural elements detection.⁴³

As observed from Table 6, research projects on CNN/GAN-based AI applications in the architectural design process were conducted between 2019 and 2024; 75% of them focused on the generation of inspiration 2D images or 3D models. Three of the twenty used AIs to generate design drawings such as floorplan and structural plan, and two studies focused on image-recognition capability to act as design analysers. Special attention is placed on several studies that integrated regression-based or RNN/transformer-based AIs to support the image processing functions of CNN/GAN-based AIs as a holistic workflow. This raises awareness of potential integrative application of different AIs rather than targeting the “best AI” with the highest speed or accuracy.

Human-AI interaction environments

In this section, the studies reveal various environments which researchers and designers interacted with CNN/GAN-based AIs, as summarised in Figure 5.

Figure 5.

Environments of human and CNN/GAN-based AI interaction from 2019 to 2024.

More than half of the AI application demonstrations were conducted on platforms with UI, showing a significant increase compared to regression-based research projects (Figure 4), where most studies were performed within code-based environments.

For example, the 3D model generation and post-processing were performed in 3D modelling software, including Rhinoceros and Grasshopper,³¹ Zbrush,³⁶ Maya⁴⁴ and Revit Dynamo.⁴⁶ On the other hand, although 2D image generation AI models were trained in code-based environments, several researchers deployed the trained image-generation models onto user-friendly web-based platforms such as RunwayML.^30,38 This allowed researchers to test the models more efficiently and enabled participants to interact with the systems intuitively without engaging with the underlying code. Similarly, one-fourth of the studies were performed solely on UI-equipped applications, especially in 2024, with more focus on the design applications, which include 3D modelling software such as Revit Dynamo,⁴² web applications,³⁹ a self-developed application on the website,³³ smartphone⁴³ and a VR device.³⁴

Notably, two of the eight studies conducted primarily in code-based environments incorporated natural language processing (NLP), allowing researchers and designers to interact with the AI systems with minimal coding.²² This suggests that NLP can ease the cognitive load associated with programming in ill-defined, design-oriented research scenarios²² by enabling a smoother switch between hard-coding and natural-language–based AI tool preparation and application.³²

To conclude, CNN/GAN-based AI research projects conducted between 2019 and 2024 displayed an increasing trend of expanding human-AI interactions from the primarily code-based environment to NLP or existing platforms with UI or even self-developed applications. This trend shows the importance of lowering the learning barrier for researchers, designers and users to interact with AI technologies with a minimum learning curve.

Evaluation methods

Only four research evaluations were solely based on AI model accuracy, comparing the AI predictions to ground truths,^42,47 the engineer’s solution,⁴¹ performance efficiency.³⁵ Qian et al., and Zhang et al.^29,33 took an additional step by demonstrating the visual outputs after testing the AI model accuracy, and one of them performed additional ablation studies and surveys for users’ feedback.³⁴ This demonstrated the effort to transition the technical research of AI training methodology to real-world design implications. Two researchers who developed their own AI platform focused on demonstrating the use of the platform and displaying visual outputs as a proof of concept.^22,45

The majority (fourteen of the twenty remaining research papers) placed the focus on demonstrating the AI-enhanced design process as a proof-of-concept. For example, del Campo³¹ and Horvath & Pouliou³² demonstrated the proposed workflows in competition entries. Eisenstadt,⁴⁰ Liu et al.³⁹ Del Campo et al.³⁶ and Sabah et al.,⁴³ demonstrated the workflow by presenting process images visually and describing them verbally. Pouliou et al.⁴⁶ took a step further than presenting the AI-enhanced design outputs to 34 practitioners for opinions, and Danchenko³⁰ reflected and iterated an improved human-AI design framework after observing the demonstration results.

More importantly, some researchers involved external parties to validate the workflows. For instance, Bank³⁸ tested the proposed framework in a course for a semester with 60 students for a house design, followed by observation of three different selected students’ approaches. Yousif & Bolojan³⁷ held a workshop for 24 students, divided into six groups, followed by applying entropy measurement of the linkography created from design process observations.

Generative AIs (RNN/transformer-based)

RNN/transformer model is responsible for encoding and decoding data sequences, and allows receiving and outputting sequences containing text, audio, and images.^48,49 Therefore, RNN/transformer-based AI models allow multi-modal applications such as Stable Diffusion, which allows text-to-image generation and image-to-image generation, or Generative Pretrained Transformers (GPT), which allows verbal discussion with image attachment. Please note that in this article, GPT refers to the AI model architecture, which is not limited to ChatGPT, which is one of the many examples of AI applications based on the GPT architecture.

The applications are categorised into three main groups, as summarised in Table 7: AI as a design assistant, an inspiration image generator and a design drawing generator. Design assistant refers to AI providing explicit advice, rather than visual generations, for the designers to be inspired. Back in 2015, Ferguson et al.⁵⁰ integrated NLP techniques with a knowledge-based system that, by evaluating the design models, the AI system could suggest materials, geometry and performances at the early design stage. Sabah et al.⁴³ applied more advanced NLP models such as GPT to provide verbal design advice to designers, while Homayounirad⁵¹ took a step further and applied BERTopic to provide design advice among AI agents and designers as a collaborative (“deliberation”) scenario. Zhong et al.⁵² applied GPT to discuss design strategies and provide prompts for inspiration image generations. Midjourney and Stable Diffusion with ControlNets are then used to generate images with the same massing image and AI-generated prompts. Cheung et al.⁵³ applied GPT as a design partner through verbal and non-verbal conversation, by collecting 3D modelling commands, design models metadata and changes in the user’s attention level, to analyse the user design behaviour and providing verbal feedback displayed on Rhinoceros screen and non-verbal feedback (changes of projection lighting colours) in a physical environment to suggest regulation of designers’ mood and behaviour.

Table 7.

Summary of RNN/transformer-based AI applications.

Category	Publication	AI technologies	Applications
AI as a design assistant	Feguson et al., 2015⁵⁰	NLP with a knowledge-based system (did not specify which AI model)	Provide verbal design advice
	Sabah et al., 2024⁴³	GPT
	Homayounirad, 2023⁵¹	BERTopic, LLM (did not specify which AI model)
	Cheung et al., 2023⁵³	GPT	Provide verbal design advice with environmental changes
	Zhong et al., 2024⁵²	GPT, Midjourney and Stable Diffusion (with ControlNets)	LLM to discuss design strategies and provide prompts for AI image generation
AI as an inspiration image generator	Ibarrola et al., 2024⁵⁹	SketchRNN	Train model for design image iterations
	B. Zhang et al., 2024⁵⁷	Llama2 and Stable Diffusion	LLM to revise prompts for AI image generation
	Shi et al., 2024⁵⁸	GPT and Stable Diffusion (with ControlNets)	LLM to revise prompts for AI image generation
	C. Zhang et al., 2023⁶³	DreamStudio2	Image generations (text-to-image and image-to-image)
	Fitriawijaya & Jeng, 2024⁶⁴	Midjourney
	Tan & Luhrs, 2024⁶⁵	Midjourney
	Maksoud et al., 2024⁶⁶	Midjourney
	Almaz et al., 2024⁶⁷	Midjourney
	Guridi et al., 2024⁶⁸	Stable Diffusion
	Yin, 2023⁶²	Amazon Transcribe, DALL-E
	Dortheimer et al., 2023⁶⁹	Midjourney, DALLE, Stable Diffusion
	Paananen et al., 2024⁷⁰	Midjourney, DALLE, Stable Diffusion
	Agkathidis et al., 2024⁷¹	Midjourney, DALLE, Stable Diffusion
	Ponzio et al., 2023⁵⁵	DALLE, Stable Diffusion	Image generations (text-to-image and image-to-image) and iterations (with ControlNets)
	Y. Zhang & Yin, 2024⁵⁶	Midjourney, Stable Diffusion
	Andreou, 2023⁵⁴	Stable Diffusion
AI as a design drawing generator	Qin et al., 2024⁶¹	ChatGPT	Translate designers’ verbal descriptions into executable computer code for structural design.
	Y. Zhang et al., 2024³³	GATConv	Transform the social network into a spatial graph
	Gaier et al., 2024⁶⁰	GPT	Translate designers’ verbal descriptions into executable computer code for layout design.

The majority of the transformer-based research papers (fifteen of 23) are categorised under the second application: AI as an inspiration image generator. Thirteen research projects used one or more AI models for text-to-image and image-to-image generation, such as Midjourney, DALLE and Stable Diffusion, to produce inspirational design images. Three studies incorporated an additional step by applying Stable Diffusion with the ControlNet plugin, which enables image generation while preserving selected features of the designer’s design image (e.g. edge detection), ensuring that the output follows the geometry of the massing model.^54–56 B. Zhang et al.⁵⁷ and Shi et al.⁵⁸ also used LLM, including Llama and GPT, to refine prompts for image generation with the mentioned tools. Despite the mentioned AI tools being pre-trained for immediate use, Ibarrola et al.⁵⁹ trained a SketchRNN model for specific early-stage design sketch use.

Finally, AI was also applied as a design drawing generator. For example, Gaier et al.⁶⁰ and Qin et al.⁶¹ used GPT to translate designers’ verbal descriptions into executable computer code for structural design and layout design drawings, respectively; and Y. Zhang et al.³³ used GATConv to transform the social network to a spatial graph, as a reference for floorplan design.

As observed on Table 7, with one exception of a research project in 2015, all RNN/transformer-based research papers were published between 2023 and 2024, potentially due to the rise of pre-trained LLM and AI image generators, allowing immediate use after 2022. Combinational applications among LLM and different AI image generators are observed, especially five research papers that used AI as a design assistant, allowing LLM to discuss and provide explicit design advice, which gradually transformed the role of AI from a prediction/generation tool to an evaluation and discussion design partner. It is worth noting that several researchers proposed how the AI application could be streamlined as part of 3D design modelling⁵⁵ or parametric design workflow.⁶² It is speculated that due to the user-friendliness of pre-trained LLM and AI image generators, there was less cognitive load for the researchers; therefore, they could focus more on the AI’s implications in the design process.

Human-AI interaction environments

The interaction environments with RNN/transformer-based AIs, which were focused between 2023 and 2024, with one exception in 2015, was summarised in Figure 6. Seventeen of the 20 research studies were performed on platforms with UI, which showed a prominent increase compared to regression-based (Figure 4) or CNN/GAN-based AIs (Figure 5). Three studies integrated the AI application into 3D modelling software, such as the Revit plugin⁵⁰ and part of the Grasshopper script in Rhinoceros.^53,71 Four researchers developed their platforms for users’ intuitive applications, such as web applications^33,57,62 and drawing platforms.⁵⁹ Eleven studies were conducted on existing web applications, including Midjourney, ChatGPT, Stable Diffusion, DALL-E, and DreamStudio. Only the remaining three studies were conducted in code-based environments, which involved the training or deployment of AI models that were not pre-trained, such as GPT or Stable Diffusion.

Figure 6.

Environments of human and CNN/GAN-based AI interaction from 2019 to 2024.

In conclusion, it is observed that for RNN/transformer-based AIs, the human-AI interaction environments are mostly UI-based, which allow immediate and intuitive applications for experimentation by researchers and designers. The observed trend of embedding necessary AIs into self-developed platforms, including web applications and portable devices (e.g., tablets), continued in 2024.

Evaluation methods

Only one study was evaluated based on AI model accuracy. Y. Zhang & Yin⁵⁶ described an accuracy evaluation method for the trained LoAR model, followed by observing designers’ image generations. Instead of accuracy, efficiency comparisons were made between the proposed AI workflow efficiency and the traditional methods.^50,61 Most research focused on demonstrating the proposed AI-enhanced workflow. Nine researchers presented the design process image visually and described verbally.^{8,52,55,60,62,64,65,67,68} Some studies went beyond demonstrations by involving participants to test with the AI systems. For instance, B. Zhang et al.⁵⁷ shared the developed web application for 2 months, and one thousand seven hundred sessions were recorded. C. Zhang et al.⁶³ evaluated further with a qualitative evaluation. After they tested the proposed workflow with eleven architecture graduate students with a median work experience of 2 years, they conducted interviews for designers’ feedback and analysis together with observational data collected during the design task.

Several researchers included quantitative evaluations after demonstration and observation. For example, Homayounirad⁵¹ counted usable outputs from the experiment participants. Agkathidis et al. conducted surveys with 18 student participants who used AI tools in the proposed framework for a semester in a design studio.⁷¹ They compared the marks of their design projects with those of studios that did not apply the AI-enhanced framework. Dortheimer et al.’s⁶⁹ data collection included screen recording during the participants’ design processes, then the percentage of different AI tools usage and the percentage of successful image generation rate were measured. Maksoud et al.⁶⁶ first listed design enquiry criteria, then evaluated scores by measuring the success rate of the AI image generations against the criteria. On the other hand, Andreou⁵⁴ generated a new design model by referencing AI-generated images, and then compared the energy performance of the original design with the AI-informed version. Alternatively, Ibarrola et al.⁵⁹ evaluated the 12 participants’ design results based on “output diversity”. Finally, Paananen et al. conducted three design sessions where five to six students worked on the same design task individually for about one and a half hours. They then evaluated using a preset Creativity Support Index (CSI).⁷⁰

In conclusion, with three exceptions of research evaluation based on AI accuracy or efficiency compared to traditional design workflows, the remaining 17 research papers focused on design process demonstration and observations, including six applied quantitative measures based on the design process and design outputs. Overall, it shows a significant shift from accuracy-based to design process observation and design output-based evaluation, compared to regression-based and CNN/GAN-based AI applications.

Findings

Findings are presented according to the three research questions introduced in the Introduction Section. The following sub-sections first synthesises how different AI technologies are applied in early-stage architectural design processes (response to RQ1), and focuses on the designer-AI interaction environments (response to RQ2). Then, it analyses how AI-enhanced design processes and design outputs are evaluated (response to RQ3), followed by a proposed framework for future research on the AI-enhanced conversational architectural design process.

AI as a conversational, agentic design partner

This section responds to the first question, “How do we utilise AI technology in the conversational architectural design process?”. From the reviewed studies, it is concluded in Figure 7 that the three types of AI technologies support six design applications, including (1) surrogate model, (2) design analysis, (3) image generation, (4) model generation, (5) design drawing generation and (6) design assistant. Due to its real-time processing capability, ANN/regression-based AIs were mainly used as surrogate models to allow traditional post-design performance simulation to become part of the design process. They were also explored for design analysis due to their numeric prediction attributes. Despite CNN/GAN-based AIs being explored for design analysis and design drawing generators due to their image processing capability, they were primarily used for inspiration images and model generations. Also, the combined applications of different AI tools as a workflow became more common, rather than experimenting to find the “best AI”. This phenomenon was even more dominant for RNN/transformer-based AIs due to their multimodal data processing and the rise of user-friendly AI chatbots and image generation applications such as ChatGPT and Stable Diffusion.

Figure 7.

AI technology categories and applications.

It is noteworthy that since the RNN/transformer-based AIs were explored, five research projects explored AI as a “design assistant” that the AI could have iterative, multimodal dialogues with designers. These studies described the role of AI as a “design partner”, referencing Gordon Pask’s Conversation Theory, rather than a passive question-answering machine,⁵³ a “mediator”, which “provides an informed choice from acceptable options”,⁵⁴ and a “negotiation and coordinator”.²⁰ They also identified designers and AI agents as a collaborative “deliberative assistant system”⁵¹ for “human-machine collaboration”³⁸ and the human-AI interaction targeted “novelty over quality (high resolution)”.⁵⁹ All the mentioned descriptions involved a mutual learning process between designers and AI, aiming to reach an agreement, which can be concluded as “conversational” properties, as coined by Gordon Pask.⁵³ To achieve such a vision, the agentic AI technique, which enables AI to plan how to respond and use tools strategically, is a potential candidate,^37,61 leveraging the attention mechanism within the transformer-based AIs, such as LLMs.⁶⁰ However, it is underexplored in empirical research at the moment.

Intuitive and collaborative platforms for stakeholders

This section responds to the second research question, “Through which interaction environments do designers interact with AI during the design process?” The trend of interaction environment used for human-AI interaction is shown in Figure 7.

Despite the fact that coding environments are inevitable for training specialised AI models, there is a steady trend of using platforms with UI for AI model deployments. With the emergence of pre-trained AI chatbots and image-generation tools such as ChatGPT and Midjourney, an increasing number of studies have shifted from code-based experimentation to using user-friendly, web-based applications since 2023. Similarly, the number of studies performed on self-developed applications, such as web or tablet applications, for intuitive experimentation increased with similar patterns. The observed trend indicates that researchers are increasingly prioritising user acceptance and usability in AI studies and experiments. For AI training stages, where working within coding environment is unavoidable, Ayman et al. stated the potential improvement by integrating LLM to enhance code-intensive processes.¹⁹ For research projects with code-based AI deployment, the potential integrations in BIM³⁰ or 3D modelling software were demonstrated.¹¹

From a technical perspective, Horvath & Pouliou stated the importance of awareness of the switch between coding language and application (natural) language of the AI design process, which reinforced the importance of exploring agentic AIs, where LLM could act as a “controller” within the research and design process,³² echoed Gaier et al.’s and Qin et al.’s studies that translates verbal descriptions into executable computer code for design actions.^60,61

Finally, the evidence suggests that user-friendly platforms foster inclusive participation in design process experiments, involving designers,⁶² consultants^51,62 and potential end-users.³³ Although this inclusivity enhances the real-world implications, it remains underexplored in the presented studies. Accordingly, there is a pressing need to develop and examine intuitive and collaborative AI-enhanced design processes.

Design process and output-based evaluation

This section responds to the third question, “How have existing studies evaluated the applications of AI in architectural design?”. The trend of evaluation methods is shown in Figure 8. Echoing the human-AI interaction environment trend, where more platforms with UI were applied for design experimentations, there is a prominent shift of AI-enhanced design process evaluation focus from AI prediction accuracy evaluation to demonstration and observation. Since 2021, the number of studies conducting workshops, followed by post-workshop quantitative and qualitative analysis, has increased drastically.

Figure 8.

The trend of human-AI interaction environments.

Therefore, instead of evaluating the proposed AI technologies themselves, the evaluation focuses on the design process framework and design outputs. Rather than aiming for AI prediction accuracy, Dortheimer et al.⁶⁹ stated the proposed AI shall be “reducing creative fixation” and “better understanding human intent”, and quantitative evaluation based on design outputs, such as measuring “output diversity”⁵⁹ and Creativity Support Index (CSI)⁷⁰ were presented. Other quantitative methods included designers’ surveys by scoring the design experience⁶³ and measurements of the usable rate of AI outputs.^51,66,69 Extending to the aspect of design thinking, researchers also evaluated AI-enhanced design processes by analysing different modes of human-AI design collaboration according to design process observations, such as convergent or divergent design thinking (Figure 9).^55,65

Figure 9.

Human-AI design evaluation trend.

Key principles for an AI-enhanced architectural design process

After analysing the research papers above, key principles for an AI-enhanced architectural design process framework are concluded in three aspects: (1) Vision, (2) Environment and (3) Evaluation.

It is envisioned that AI acts as a conversational design partner that allows a continuous and iterative design discussion with the designers, whereas LLM acts as the key communicator with the designers due to its natural language processing capabilities, which enables immediate communication with users with no technical barrier. Multimodal AI is advised to be integrated into architectural design processes, such as image recognition for design analysis and image generation for design iterations. To enhance the human-AI iterative dialogues, agentic AI techniques shall be investigated, which extend AI’s capabilities from providing specific functions to being flexible to adapt to different design enquiries and scenarios.

The human-AI interaction environment shall be developed on an intuitive, collaborative platform that minimises the learning curve for users to navigate the platform. For one designer scenario, AI can be integrated into the specific 3D modelling software for seamless communication and model iteration within the modelling environment. For multiple designers’ or stakeholders’ collaboration, AI applications can be deployed on web-based or self-developed platforms, such as a group messaging environment, which minimises the cognitive load for users to navigate the interface. For more participatory design scenarios, such as community planning projects, on-device integration, such as mobile or tablet applications, could be easily accessible to the public.

For design process and output-based evaluation, it is suggested that quantitative evaluation methods, such as counting usable AI output percentages, can be used for analysing design outputs; the statistics could be intuitive and valuable for future development improvement or guidelines for future designers. Qualitative evaluation methods, such as surveys and interviews, are suggested to be used for design process model analysis, in which more elaborate feedback could be collected from the designers. The discussed key principles are summarised in Table 8.

Table 8.

Key principles for an AI-enhanced architectural design process framework.

Aspect	Attributes	Methodology
Vision	AI as a conversational design partner	LLM as the key communicator
		Multimodal AI for processing design information (e.g. text, images, etc.)
		Agentic AI techniques for bridging design functions (e.g. Image recognition, generation, knowledge-based search, etc.
Environment	Intuitive and collaborative platform	Software integration for designers (e.g., plugin, LLM integration, etc.)
		Web-based/self-developed platforms for potential stakeholders
		On-device integration (e.g., mobile, tablet, etc.)
Evaluation	Design process and output-based	A quantitative measure for design outputs (e.g., usable AI outputs percentage, etc.)
Evaluation	Design process and output-based	Qualitative evaluation for the design process (e.g., survey, interview, design mode observations, etc.)

Proposed framework for a conversational AI-enhanced architectural design process

According to the findings above, a framework for a conversational AI-enhanced architectural design process is proposed (Figure 10). It is envisioned that AI is positioned as a conversational design partner by leveraging the intuitive human-AI communication capability of LLM and the multi-functional adaptivity of a multimodal, multi-agent system. The environment for interaction between designers and the AI systems is primarily in group chat environments, in either instant messaging platforms such as WeChat or Discord, or web-based applications, which minimise the designers’ learning curve. To validate the proposed framework, conducting design workshops with small groups of designers is advised. To evaluate and improve the proposed system, a hybrid evaluation method is proposed. Quantitative methods include pre-workshop and post-workshop surveys and counts of usable human-AI interactions, while qualitative measures include semi-structured interviews to assess designers’ feedback.

Figure 10.

Proposed AI-enhanced conversational architectural design process evaluation framework.

Discussion and limitations

This systematic review addresses the limitations of current reviews in the field of AI-enhanced architectural design process, where they lack design process focus, exclusive coverage of AI technologies and under-exploration of the environments for human-AI interaction. This proposed key principle for an AI-enhanced architectural design framework highlights the vision of AI as a conversational design partner, facilitating iterative discussions with designers through natural language processing. Integrating multimodal AI, including image recognition and generation, is recommended for design analysis and iterations. Agentic AI techniques should be explored to enhance flexibility in responding to diverse design inquiries. The interaction environment should be an intuitive, collaborative platform to reduce the learning curve. In single-designer scenarios, AI can be embedded in specific 3D modelling software for seamless communication. For collaborative efforts among multiple designers, web-based platforms or group messaging environments are ideal. Community planning initiatives could utilise accessible mobile or tablet applications. For evaluation, quantitative methods, such as tracking usable AI output percentages, can provide insights for improvement. Qualitative methods like surveys and interviews can capture detailed feedback on the design process and outputs.

Implications to research, design practice and education

Amidst the rapid emergence of AI technologies, this study provides an overview of their applications in the architectural design process. The identified key principles offer guidance for researchers to develop AI-enhanced design frameworks and tools. Positioned as an efficient “cookbook” for researchers, this work clarifies the suitability of ANN-based AIs for training surrogate models that generate immediate performative predictions. It further advocates for integrating AI within 3D modelling software, where LLMs facilitate intuitive, natural language-based communication between designers and AI systems. For evaluation, on top of the accuracy comparison between the AI prediction and simulation results, the researcher could analyse the conversational history between the designer and the AI bot and prepare a post-workshop survey and interview for designers’ feedback as a holistic evaluation framework. Similarly, the evaluation methods could benefit architectural design education. Since it is becoming more common for architecture courses to experiment with integration with AI tools, tutors could prepare a collaborative design discussion room where tutors and students can converse with AI bots in a group messaging environment. The AI bot could be equipped with a knowledge base about the design brief. Evaluation could be focused on assessing divergent and convergent thinking based on the design process throughout different stages of the design project. These two examples demonstrate the potential implications of the key principles in current situations, avoiding cognitive overload for researchers and tutors to prepare for an AI-enhanced architectural design process framework with the exploding development of AI technologies.

Design process and output-based evaluation

Although the study provides an overall view of the AI-enhanced architectural design process, several limitations exist. Since AI technology is developing rapidly, the latest AI applications in the architecture field might be neglected in this study. Although the proposed principles primarily focus on the mode of the human-AI collaborative design process, revolutionised AI applications might impact the discussed principles. For instance, ChatGPT and Midjourney caused a dominant shift in AI-enhanced design processes since 2021. Also, ethical issues such as data bias, potential copyright and plagiarism are not explored. They are crucial to be taken into consideration before real-world practice implementation. Future research studies could be conducted by carefully considering the ethical challenges, followed by applying the proposed principles for an empirical study with a developed multimodal, agentic AI tool for design collaboration.

Conclusion

This paper reports on a systematic review process that, from an initial pool of 1138 results from 8 databases between 2014 and 2024, resulted in an analysis of 63 papers. They are categorised into three categories according to the types of AI technologies, including regression-based, CNN/GAN-based and RNN/transformer-based. Each category is further studied according to three aspects: (1) AI applications in the design process, (2) the environment of human-AI interaction, and (3) design process evaluation. After discussing the trends and limitations of the current research, three key principles for the AI-enhanced architectural design process are established. The role of AI in architectural design is anticipated to be that of a conversational design partner. The ways in which humans and AI interact should be collaborative and intuitive, with evaluations focused on the design process and its outcomes. Then, a novel, comprehensive framework for a conversational AI-enhanced design process is proposed, including vision, human-AI interaction environments and evaluation methods. Finally, the research implications are discussed alongside the identified limitations and future research opportunities.

Footnotes

ORCID iDs

Lok Hang Cheung

Asterios Agkathidis

Juan Carlos Dall’Asta

Giancarlo Di Marco

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Xi’an Jiaotong-Liverpool University under XJTLU Postgraduate Research Scholarship (PGRSB211206).

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

Data sharing is not applicable to this article as no new data were created or analysed in this study.*

References

Castro Pena

Carballal

Rodríguez-Fernández

, et al. Artificial intelligence applied to conceptual design. A review of its use in architecture. Autom Constr 2021; 124: 103550.

Momade

Durdyev

Estrella

, et al. Systematic review of application of artificial intelligence tools in architectural, engineering and construction. Front Eng Built Environ 2021; 1: 203–216.

Saka

Oyedele

Akanbi

, et al. Conversational artificial intelligence in the AEC industry: a review of present status, challenges and opportunities. Adv Eng Inform 2023; 55: 101869.

Rane

Choudhary

Rane

. Leading-edge technologies for architectural design: a comprehensive review. SSRN Journal 2023.

Enjellina

BEVP

Rossy

AGC

. A review of AI image generator: influences, challenges, and future prospects for architectural field. J Artif Intell Archit 2023; 2: 53–65.

Zhang

, et al. Generative AI models for different steps in architectural design: a literature review. Front Archit Res 2024; 14: 759–783.

Vissers-Similon

Dounas

De Walsche

. Classification of artificial intelligence techniques for early architectural design stages. Int J Archit Comput 2024; 23: 387–404.

Cheung

Dall’Asta

. Exploring a collaborative and intuitive framework for combined application of AI art generation tools in architectural design process. Proceedings of the AISB Convention 2023, 2023, pp. 122–130.

Deep learning specialization. Coursera. https://www.coursera.org/specializations/deep-learning/ (accessed 16 January 2025).

10.

Machine learning specialization. Coursera. https://www.coursera.org/specializations/machine-learning-introduction/ (accessed 16 January 2025).

11.

Radziszewski

Waczyńska

. Machine learning algorithm-based tool and digital framework for substituting daylight simulations in early-stage architectural design evaluation. Proceedings of the Symposium on Simulation for Architecture and Urban Design. Society for Computer Simulation International, 2018, pp. 1–7.

12.

Sebestyen

Tyc

. Machine learning methods in energy simulations for architects and designers - the implementation of supervised machine learning in the context of the computational design process. In: Anthropologic: Architecture and Fabrication in the cognitive age - Proceedings of the 38th eCAADe Conference - Volume 1, Werner

Koering

(eds), Berlin, Germany, 16-18 September 2020, TU Berlin, 2020, pp. 613–622.

13.

Lin

C-H

Tsay

Y-S

. A metamodel based on intermediary features for daylight performance prediction of façade design. Build Environ 2021; 206: 108371.

14.

Hasan

Horvat

. An application of the random forest algorithm for the prediction of solar envelope ‘Floor Space Index’ based on spatiotemporal parameters. J Build Eng 2024; 86: 108784.

15.

Salamanca

Apolinarska

Pérez-Cruz

, et al. Augmented intelligence for architectural design with conditional autoencoders: Semiramis case study. In: Gengnagel

Baverel

Betti

, et al. (eds). Towards Radical Regeneration. Springer International Publishing, 2023, pp. 108–121.

16.

Rahmani Asl

. Energy model machine (EMM) - instant building energy prediction using machine learning. In: ShoCK! - Sharing Computational Knowledge! - Proceedings of the 35th eCAADe Conference - Volume 2, Fioravanti

Cursi

Elahmar

, et al. (eds), Rome, Italy, 20-22 September 2017, Sapienza University of Rome, 2017, pp. 277–286.

17.

Wang

Shi

, et al. Automatic design of main pedestrian entrance of building site based on machine learning - a case study of museums in China’s Urban environment. In: Learning, Adapting and Prototyping - Proceedings of the 23rd CAADRIA Conference - Volume 2, Fukuda

Huang

Janssen

, et al. (eds), Beijing, China, 17–19 May 2018, Tsinghua University, 2018, pp. 227–235.

18.

P-K

Liu

S-Y

. The application of artificial neural networks to facilitate the architectural design process. Adv Intell Syst Comput 2020; 1069: 79–89.

19.

Ayman

Mansour

Eldaly

. Applying machine learning algorithms to architectural parameters for form generation. Autom Constr 2024; 166: 105624.

20.

Gerber

Pantazis

Marcolino

, et al. A multi agent systems for design simulation framework: experiments with virtual physical social feedback for architecture. In: Proceedings of the Symposium on Simulation for Architecture & Urban Design, San Diego, CA, Society for Computer Simulation International, 2015, pp. 205–212.

21.

Wang

Snooks

. Intuitive behavior - the operation of reinforcement learning in generative design processes. In: PROJECTIONS - Proceedings of the 26th CAADRIA Conference - Volume 1, Globa

van Ameijde

Fingrut

, et al. (eds), Hong Kong, 29 March–1 April 2021, The Chinese University of Hong Kong and Online, 2021, pp. 101–110.

22.

Başarır

Erol

. Briefing AI: from architectural design brief texts to architectural design sketches. In: Architecture in the Age of Disruptive Technologies: Transformations and Challenges [9th ASCAAD Conference Proceedings], Abdelmohsen

El-Khouly

Mallasi

, et al. (eds), Cairo (Egypt), 2021, pp. 23–31.

23.

Algeciras-Rodríguez

. Trained architectonics. In: Parametricism Vs. Materialism: Evolution of Digital Technologies for Development [8th ASCAAD Conference Proceedings ISBN 978-0-9955691-0-2], London, 7-8 November 2016, CUMINCA, 2016, pp. 461–468.

24.

Harding

. Dimensionality reduction for parametric design exploration Conference or Workshop Item. In: Adriaenssens

Gramazio

Kohler

, et al. (eds). Advances in Architectural Geometry, 2018.

25.

Kazemi

Entezami

Ghisi

. Machine learning techniques for diagrid building design: Architectural–Structural correlations with feature selection and data augmentation. J Build Eng 2024; 86: 108766.

26.

Kim

Johanes

Huang

. Flow2Form: a flow-driven computational framework for early stage architectural design. In: Proceedings of the 43rd Annual Conference for the Association for Computer Aided Design in Architecture (ACADIA). Crawford

Diniz

Beckett

, et al. (eds). 2023, pp. 408–418. ACADIA.

27.

Millán

Belmonte

M-V

Boned

F-J

, et al. Using machine learning techniques for architectural design tracking: an experimental study of the design of a shelter. J Build Eng 2022; 51: 104223.

28.

Schaumann

Kalay

Hong

, et al. Simulating human behavior in not-yet built environments by means of event-based narratives. Proceedings of the Symposium on Simulation for Architecture & Urban Design. Society for Computer Simulation International, 2015, pp. 5–12.

29.

Qian

Yang

Mei

, et al. Artificial intelligence-designer for high-rise building sketches with user preferences. Eng Struct 2023; 275: 115171.

30.

Danchenko

. The AI-teration method and the role of AI in architectural design. In: Arai

Kapoor

Bhatia

(eds). Proceedings of the Future Technologies Conference (FTC) 2020. Springer International Publishing, 2021, vol 1, pp. 525–538.

31.

del Campo

. Architecture, Language and AI - language, attentional generative adversarial networks (AttnGAN) and Architecture design. In: PROJECTIONS - Proceedings of the 26th CAADRIA Conference, Globa

van Ameijde

Fingrut

, et al. (eds), 29 March - 1 April 2021, CUMINCAD, 2021, pp. 211–220. The Chinese University of Hong Kong and Online.

32.

Horvath

A-S

Pouliou

. AI for conceptual architecture: reflections on designing with text-to-text, text-to-image, and image-to-image generators. Front Archit Res 2024; 13: 593–612.

33.

Zhang

Peng

. Participatory space Auto-Encapsulation: bridging human relationships to building typology with machine intelligence. In: ACCELERATED DESIGN - Proceedings of the 29th CAADRIA Conference, Gardner

Herr

Wang

, et al. (eds), Singapore, 20-26 April 2024, 2024, pp. 231–240.

34.

Rasoulzadeh

Wimmer

Stauss

, et al. Strokes2Surface: recovering curve networks from 4D architectural design sketches. Comput Graph Forum 2024; 43: e15054.

35.

Ampanavos

Malkawi

. Early-phase performance-driven design using generative models. In: Gerber

Pantazis

Bogosian

, et al. (eds). Computer-Aided Architectural Design. Design Imperatives: The Future is Now. Springer, 2022, pp. 87–106.

36.

del Campo

Carlson

Manninger

. Towards hallucinating machines - designing with computational vision. Int J Archit Comput 2021; 19: 88–103.

37.

Yousif

Bolojan

. Strategies of interconnecting deep learning models in AI-Driven design systems. In: Di Marco

Lombardi

Tedjosaputro

(eds). Creativity in the Age of Digital Reproduction. Springer Nature, 2024, pp. 244–252.

38.

Bank

Sandor

Schinegger

, et al. Learning spatiality - a GAN method for designing architectural models through labelled sections. In: Pak

Wurzer

Stouffs

(eds). Co-creating the Future: Inclusion in and through Design - Proceedings of the 40th Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2022). Ghent, 2022, pp. 611–619.

39.

Liu

Shen

Ren

, et al. Pipes of AI – machine learning assisted 3D modeling design. In: Yuan

Yao

Yan

, et al. (eds). Proceedings of the 2020 DigitalFUTURES. Springer, 2021, pp. 17–26.

40.

Eisenstadt

. Generation of floor plan variations with convolutional neural networks and Case-based reasoning - an approach for transformative adaptation of room configurations within a framework for support of early conceptual design phases. In: Architecture in the Age of the 4th Industrial Revolution - Proceedings of the 37th eCAADe and 23rd SIGraDi Conference, Sousa

Xavier

Castro Henriques

(eds). University of Porto, Porto, Portugal, 2019, pp. 79–84.

41.

Liao

Huang

, et al. Automated structural design of shear wall residential buildings using generative adversarial networks. Autom Constr 2021; 132: 103931.

42.

Clever

Abualdenien

Borrmann

. Deep learning approach for predicting pedestrian dynamics for transportation hubs in early design phases. 2021.

43.

Sabah

Alani

Salih

BMM

. Leveraging AI language models for designing contextually responsive built environments. Lect Notes Netw Syst 2024; 1019 LNNS: 510–519.

44.

del

Carlson

Manninger

. 3D graph convolutional neural networks in architecture design. In: Proceedings of the 40th Annual Conference of the Association of Computer Aided Design in Architecture (ACADIA), Slocum

Ago

Doyle

, et al. (eds), 24–30 October 2020, Online and Global, 2020, pp. 688–696.

45.

Ennemoser

Mayrhofer-Hufnagl

. Design across multi-scale datasets by developing a novel approach to 3DGANs. Int J Archit Comput 2023; 21: 358–373.

46.

Pouliou

Horvath

A-S

Palamas

. Speculative hybrids: investigating the generation of conceptual architectural forms through the use of 3D generative adversarial networks. Int J Archit Comput 2023; 21: 315–336.

47.

Zhuang

Yang

, et al. Synthesis and generation for 3D architecture volume with generative modeling. Int J Archit Comput 2023; 21: 297–314.

48.

Cho

van Merriënboer

Gulcehre

, et al. Learning phrase representations using RNN encoder–decoder for statistical machine translation. In: Moschitti

Pang

Daelemans

(eds). Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, pp. 1724–1734.

49.

Sutskever

Vinyals

. Sequence to sequence learning with neural networks. Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2. Cambridge, MA, USA: MIT Press, 2014, pp. 3104–3112.

50.

Ferguson

Vardeman

Buccellato

APC

. Capturing an architectural knowledge base utilizing rules engine integration for energy and environmental simulations. Proceedings of the Symposium on Simulation for Architecture & Urban Design. Society for Computer Simulation International, 2015, pp. 67–74.

51.

Homayounirad

. Designing the built environment through hybrid intelligence. HHAI 2023: Augmenting Human Intellect. IOS Press, pp. 483–495.

52.

Zhong

Shi

Cheung

, et al. AI-Enhanced performative building design optimization and exploration: a design framework combining computational design optimization and generative AI. In: ACCELERATED DESIGN - Proceedings of the 29th CAADRIA Conference. Singapore, 2024, pp. 59–68.

53.

Cheung

Dall’Asta

Di Marco

. Exploring large language model as a design partner through verbal and non-verbal conversation in architectural design process. XXVII International Conference of the Ibero-American Society of Digital Graphics. Blucher Proceedings, pp. 1059–1070.

54.

Andreou

Kontovourkis

Solomou

, et al. Rethinking architectural design process using integrated parametric design and machine learning principles. In: Dokonal

Hirschberg

Wurzer

(eds). Digital Design Reconsidered - Proceedings of the 41st Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2023). Graz, 2023, pp. 461–470.

55.

Ponzio

Chornobai

Vinicius

, et al. Exploring creative AI thinking in the design process: the design intelligence strategy. In: HUMAN-CENTRIC - Proceedings of the 28th CAADRIA Conference, Ahmedabad, Koh

Reinhardt

Makki

, et al. (eds), 2023, pp. 49–58.

56.

Zhang

Yin

. Application of AIGC technology in vernacular architecture design. In: 2024 4th International Conference on Consumer Electronics and Computer Engineering (ICCECE), pp. 492–497.

57.

Zhang

Biao

, et al. SIMForms: a web-based generative application fusing forms, metrics, and visuals for early-stage design. In: ACCELERATED DESIGN - Proceedings of the 29th CAADRIA Conference, Singapore, Gardner

Herr

Wang

, et al. (eds), 2024, pp. 373–382.

58.

Shi

Seo

Cha

, et al. Generative AI-powered architectural exterior conceptual design based on the design intent. J Comput Des Eng 2024; 11: 125–142.

59.

Ibarrola

Lawton

Grace

. A collaborative, interactive and context-aware drawing agent for Co-Creative design. IEEE Trans Vis Comput Graph 2024; 30: 5525–5537.

60.

Gaier

Stoddart

Villaggi

, et al. Generative design through quality-diversity data synthesis and language models. In: Proceedings of the Genetic and Evolutionary Computation Conference, New York, NY, Association for Computing Machinery, pp. 823–831.

61.

Qin

Guan

Liao

, et al. Intelligent design and optimization system for shear wall structures based on large language models and generative artificial intelligence. J Build Eng 2024; 95: 109996.

62.

Yin

. Speedy façade. In: Dokonal

Hirschberg

Wurzer

(eds). Digital Design Reconsidered - Proceedings of the 41st Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2023). Graz, 2023, pp. 409–418.

63.

Zhang

Wang

Pangaro

, et al. Generative image AI using design sketches as input: opportunities and challenges. Proceedings of the 15th Conference on Creativity and Cognition. Association for Computing Machinery, pp. 254–261.

64.

Fitriawijaya

Jeng

. Integrating multimodal generative AI and blockchain for enhancing generative design in the early phase of architectural design process. Buildings 2024; 14: 2533.

65.

Tan

Luhrs

. Using generative AI midjourney to enhance divergent and convergent thinking in an architect’s creative design process. Des J 2024; 27: 677–699.

66.

Maksoud

Elshabshiri

Saeed Hilal Humaid Alzaabi

, et al. Integrating an image-generative tool on creative design brainstorming process of a safavid mosque architecture conceptual form. Buildings 2024; 14: 843.

67.

Almaz

El-Agouz

EAE

Abdelfatah

, et al. The future role of artificial Intelligence (AI) design’s integration into architectural and interior design education is to improve efficiency, sustainability, and creativity. Civ Eng Archit 2024; 12: 1749–1772.

68.

Guridi

Cheyre

Goula

, et al. Image generative AI to design public spaces: a reflection of how AI could improve Co-Design of public parks. Digit Gov Res Pr 2024; 6: 1–14.

69.

Dortheimer

Schubert

Dalach

, et al. Think AI-side the box! exploring the usability of Text-to-Image generators for architecture students. Digital Design Reconsidered - eCAADe 2023, 2023, pp. 567–576.

70.

Paananen

Oppenlaender

Visuri

. Using text-to-image generation for architectural design ideation. Int J Archit Comput 2024; 22: 458–474.

71.

Agkathidis

Yang

Symeonidou

. AI-Assisted design: utilising artificial intelligence as a generative form-finding tool in architectural design studio teaching. In: Data-Driven Intelligence - Proceedings of the 42nd Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2024). Nicosia, Kontovourkis

Phocas

Wurzer

(eds). 2024, pp. 619–628.