Automation in Construction 159 (2024) 105264

Contents lists available at ScienceDirect

Automation in Construction
journal homepage: www.elsevier.com/locate/autcon

Intelligent floor plan design of modular high-rise residential building based

on graph-constrained generative adversarial networks
Jiepeng Liu a, b, Zijin Qiu a, b, Lufeng Wang a, b, Pengkun Liu b, d, *, Guozhong Cheng a, b, *,
Yan Chen c
a
Key Laboratory of New Technology for Construction of Cities in Mountain Area (Chongqing University), Ministry of Education, Chongqing 400045, China
b
School of Civil Engineering, Chongqing University, Chongqing 400045, China
c
China State Construction Engineering Corporation Xinjiang Construction & Engineering Corporation Ltd, Urumqi 830063, Xinjiang, China
d
Department of Civil and Environmental Engineering, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213, United States

A R T I C L E I N F O A B S T R A C T

Keywords: In the context of Modular High-Rise Residential Buildings (MHRBs), designing floor plans involves intricate
Floor plan design complexities due to the need for adherence to numerous domain-specific design rules. To address this issue, our
Modular high-rise residential buildings research introduces a novel framework based on a Graph-Constrained Generative Adversarial Network (GC-GAN)
Graph-constrained generative adversarial
specialized for generating MHRB floor plans. This enhanced GC-GAN incorporates knowledge graphs that
networks
Knowledge graph
encapsulate domain-specific constraints and guidelines, thereby generating floor plans that exhibit realism, di­
Knowledge-based design versity, and conformity to established design principles. Additionally, the framework integrates a sophisticated
image-to-vector conversion algorithm that enables seamless alignment with a predefined flat-design standardi­
zation library. A salient feature of this framework is the automated generation of Building Information Modeling
(BIM) models, which rigorously conform to the modularity specifications essential for efficient modular con­
struction. The efficacy and practical applicability of our approach have been validated through an exhaustive
analysis covering fifteen cases across five diverse scenarios.

1. Introduction and background structures. Designing MHRBs requires careful planning and structural
consideration to ensure safety and durability [1,2,5,14]. Ensuring suf­
Modular buildings are increasingly adopted globally, acclaimed for ficient thermal and acoustic insulation is also challenging, as these
their high-quality construction, efficient assembly, and commitment to buildings must meet the strict standards of high-rise construction.
environmental sustainability [1–4]. These structures require off-site Consequently, MHRBs’ floor plan design present more complex chal­
fabrication of prefabricated modules, which are then transported and lenges, requiring the navigation of numerous design requirements
assembled on-site [5]. This approach necessitates additional design [15–18]. Traditional design methods, dependent on iterative processes
considerations early on, often leading to errors in the design phase. and the expertise of designers, are not only time-consuming but also
Research shows that such errors can increase the total contract value by susceptible to inaccuracies [19,20]. Implementing automated floor plan
5–20% and represent 14.21% of that value [6,7]. The floor plan is a design in MHRBs can markedly diminish design errors and enhance
critical element in this phase [8–11], governing spatial functionality, operational efficiency, offering benefits to the Architecture, Engineer­
structural integrity, and the building’s energy efficiency [12,13]. The ing, and Construction (AEC) industry.
incorporation of modular buildings into high-rise residential construc­ Advancements in intelligent algorithms and Artificial Intelligence-
tion poses further challenges. Modular High-Rise Residential Buildings Generated Content (AIGC) offer promising solutions for automating
(MHRBs) are characterized by the use of interconnected prefabricated building layout plans, significantly enhancing design efficiency and
modules and core shafts, which encompass elements such as staircases effectiveness. In contrast to traditional architectural methods, auto­
and elevators. Their design involves significant complexities due to the mated approaches can generate multiple viable solutions for designers,
intricate connections of prefabricated core shaft modules and is further substantially enhancing design outcomes. Although prior research
complicated by increased wind pressures and seismic forces in high-rise [21–23] has explored the application of traditional evolutionary

* Corresponding authors at: School of Civil Engineering, Chongqing University, Chongqing 400045, China.
E-mail address: [email protected] (P. Liu).

https://doi.org/10.1016/j.autcon.2023.105264
Received 7 September 2023; Received in revised form 9 December 2023; Accepted 30 December 2023
Available online 16 January 2024
0926-5805/© 2024 Elsevier B.V. All rights reserved.
J. Liu et al. Automation in Construction 159 (2024) 105264

algorithms, such as growth algorithms, in the generation of room lay­ (3) proposing the use of graph-constrained GANs (GC-GAN) to generate
outs, these methods fall short of capturing and representing the intricate diverse and compliant floor plans for MHRBs, guided by a knowledge
design knowledge and complexities inherent in floor plan design. Such graph; (4) introducing a matching method that vectorizes generated
methods also struggle to address specific challenges associated with floor plans to align with templates from a standard design library,
high-rise residential buildings, like optimizing connectivity and circu­ addressing the instability issues common in traditional GAN output; (5)
lation. Consequently, these limitations often lead to unforeseen design incorporating an automatic BIM generation method to expedite the
modifications, rework, material waste, and, ultimately, additional costs modeling process.
and delays [24]. The subsequent sections are organized as follows: Section 2 reviews
Knowledge-based design methods, which amalgamate domain- the literature. Section 3 delineates the problem formulation, relevant
specific expertise with cutting-edge computational techniques, have design knowledge, and evaluation metrics for MHRBs’ architectural
emerged as efficacious alternatives to traditional architectural design floor plan design. Section 4 describes the methodology for the graph-
paradigms. Techniques leveraging graph theory and Generative Adver­ constrained GAN-based design framework, the prepared floor plan
sarial Networks (GANs) have demonstrated considerable utility across dataset, and the floor plan matching process with a standard design li­
diverse domains, including but not limited to architecture and urban brary. Section 5 presents a validation study that evaluates the generated
planning generation [13–15], as well as indoor layout generation floor plans and their conformity to the standard design library, and
[25–29]. Such methodologies empower architects to generate and crit­ Section 6 concludes with remarks and directions for future research.
ically evaluate design alternatives with enhanced efficiency and objec­
tivity, thereby optimizing both building performance and user 2. Literature review
satisfaction [16]. While some research has explored the application of
knowledge-based design methods in architecture or indoor layout gen­ 2.1. Architectural floor plan design challenges of modular high-rise
eration [25–29], there remains a conspicuous absence of targeted in­ residential building
vestigations concerning MHRBs, particularly in the delineation of
specific design requirements for these structures. Amidst increasing urbanization and population growth, the design of
The aim of this paper is to advance the automation of MHRBs’ floor MHRBs has become a focal point, necessitating the reconciliation of
plan generation in alignment with design requirements. The study diverse design preferences, quality metrics, and production parameters
tackles three crucial research questions: in the face of numerous stakeholder concerns [24,30,31].
With regard to design preferences, essential factors to consider
(1) What comprises domain-specific knowledge in MHRBs’ archi­ include the number of housing units, staircases, elevators, and corridors,
tectural floor plan design? aligned with the goals for occupancy levels. Francesca Russello Ammon
(2) How can this knowledge be accurately reflected in MHRBs’ floor [32] underscores the necessity of tailoring design to market demands for
plan representations? occupancy, especially in burgeoning urban landscapes. Compliance with
(3) Using this knowledge, how can optimized floor plan designs for both market demands and regulatory frameworks in determining unit
MHRBs be generated efficiently? sizes is crucial. Studies by Tokunaga and Murota [33] identify consumer
preferences in unit selection and size as critical determinants of a
Addressing these questions presents several technical challenges. building’s marketability. They further elaborate that the life stages of
Firstly, the complex nature of MHRBs’ floor plans necessitates further the residents significantly influence their preferences for different types
elucidation of what domain-specific knowledge and constraints should of units—such as studios, one-bedroom apartments, or penthouses.
be factored in and how they can be gleaned from existing design sche­ In terms of design quality, the interconnectedness of spaces is crucial
matics. Secondly, an effective method for extracting pertinent design to eliminating isolated regions and enhancing community living. Javid
knowledge from existing MHRB’s floor plans must be formulated. Lastly, Ghanbari [34] discusses the psychological benefits of well-connected
considerable effort is required to establish an automated, intelligent spaces. Efficient circulation between spatial units, considering path
mechanism for generating MHRBs’ floor plans predicated on this complexity and length, enhances resident mobility and comfort. Optimal
accrued design knowledge. space utilization is paramount, aiming to minimize empty or unusable
To address these challenges, this study presents an automated spaces between units. Designs should focus on ensuring that each living
framework for generating floor plans of MHRBs. This framework em­ space receives sufficient daylight. Kılıç and Yener [35] identify
ploys knowledge graphs to encapsulate and articulate design expertise, daylighting as an essential aspect of both visual comfort and energy
suggests the adoption of graph-constrained GANs for the acquisition of efficiency. Adherence to standardized dimensions for elements such as
expertise in graphic design, and introduces a planar matching algorithm windows and doors is not only aesthetically coherent but also func­
to match the outcomes of the graph-constrained GANs with existing tionally effective.
projects, facilitating the creation of BIM models. The framework in­ As for the production requirements, adhering to maximum di­
tegrates floor plan design and BIM post-processing to automate the mensions during the building process is essential, considering material
conversion of floor plan designs into 3D models. It not only supports cuts and assembly. Lu et al. [36] and Griffin [37] discuss automated
design interactivity but also maintains practicality, culminating in a structural design and the potential of cross-laminated timber, respec­
personalized and automated design process for MHRBs, improving the tively, as innovations in fabrication. A specific minimum storage space
efficiency and convenience of design. for materials should be designated to optimize workflow and reduce
This research sets three specific objectives: (1) define the design construction delays. The design should facilitate the easy transport of
requirements, constraints, and metrics for MHRBs’ floor plans from materials to, from, and within the construction site. Logistical efficiency
existing drawings; (2) introduce knowledge graphs for representing is a critical consideration in high-rise construction. By meticulously
design knowledge derived from such drawings; (3) develop and assess a addressing each of these requirements, high-rise residential building
graph-constrained GAN-based framework’s effectiveness in generating design can achieve functionality, aesthetic appeal, and production effi­
floor plan designs for MHRBs. ciency. This aligns with the overarching goal of creating structures that
In summary, this study’s contributions are as follows: (1) establish­ are not only livable but also sustainable and marketable.
ing a comprehensive framework detailing the design knowledge and The foregoing discussion underscores the significant research gaps,
requirements specific to MHRBs; (2) presenting a knowledge graph- including the absence of a well-defined knowledge base specific to
based methodology for capturing and articulating MHRBs’ design MHRBs, the imperative for an integrative approach that combines
knowledge, with clear explications of the knowledge graph components; human expertise with technology, and the necessity for advanced

2
J. Liu et al. Automation in Construction 159 (2024) 105264

models that learn from past designs. Addressing these gaps is critical to et al. [68] suggested a data-driven generative design framework using
enhancing the design of MHRBs in response to the challenges of esca­ GANs, incorporating multiple data types to enhance automation and
lating urban density and population sizes. performance in detailed design, ultimately improving design efficiency
and user satisfaction. Additionally, Huang and Zheng [50], Chaillou
2.2. Knowledge representation from design [69], and Nauata et al. [46] employed GANs to learn from architectural
design drawings and generate house layouts. Consequently, GAN tech­
Traditional methods in structural design encompass topology opti­ niques hold the potential to learn expert experience and domain
mization, genetic algorithms, cellular automata, and generative and knowledge and generate the design according to the learned knowledge
analogy design. These conventional approaches are predicated on pre­ automatic.
defined rules and cannot learn and represent design codes and con­ The above studies demonstrate the feasibility of utilizing GAN
structibility constraints [21,22]. models for generative design in combination with knowledge-based
Representing human expertise from textual sources enables effective design principles across various research fields. However, there is
knowledge transfer organization and informs decision-making [38]. currently a lack of research on floor plan designs for MHRBs and a
Extracting and organizing relevant information aids in understanding deficiency in knowledge-based generative models. This highlights the
complex issues and fostering effective communication between stake­ necessity for an integrated approach that combines knowledge-based
holders. Researchers have investigated the challenges of understanding design principles with generative models, as well as the development
and learning human knowledge through knowledge graphs [39–44], of enhanced models capable of learning design knowledge from previous
GAN [45–50], and so on [51–57]. For example, Liu et al. [58] developed design drawings. In summary, Our innovation is focused on introducing
a novel data-driven framework based on BIM and knowledge graphs to a knowledge-based generative model and a subsequent BIM generative
automatically represent unstructured BIM data and infer auditing results modeling algorithm. The main objective of this approach is to facilitate
for BIM model errors. Hao et al. [59] introduced a method for auto­ the continuous flow of information throughout the entire building cycle,
matically converting word-based compromise Decision Support Prob­ encompassing the design and production phases. Different methods are
lems into knowledge graphs for the supply chain design domain. Liu compared in Table 1.
et al. proposed an automated rebar design framework to learn and
represent the design knowledge from existing detailed rebar design 2.4. Research gaps
drawings through GAN and deep reinforcement learning [45,60].
The research mentioned above shows the potential to learn and This review identifies several research gaps in the existing literature,
represent design knowledge through the knowledge graphs and GAN. emphasizing the need for automated floor plan designs for MHRBs. First,
However, there is a lack of knowledge of representation methods for the the extant research lacks a comprehensive definition of design knowl­
MHRBs’ design to represent the design knowledge and requirements edge and specific requirements tailored to MHRBs. Second, there is an
from the existing floor plan designs. absence of methods for effectively capturing and representing design
knowledge from existing MHRB layouts. Third, while current models
2.3. Knowledge-based design often focus either on knowledge-based design or on productivity, they
rarely integrate both. Furthermore, existing models generally fall short
Knowledge-based design approaches use structured knowledge to of addressing specific design challenges unique to MHRBs. Conse­
inform and guide the design process [61]. The study by Maher et al. quently, there is a pressing need for advanced models capable of
demonstrates how integrating knowledge-based systems in architectural generating designs that not only adhere to crucial design principles but
design can facilitate the generation of creative and innovative solutions also maintain a high level of innovation and functionality. These three
[62]. Wei et al. introduce a knowledge-based quality analysis system research gaps are in line with the floor plan design challenges for MHBRs
designed explicitly for model-based definition (MBD) part models using presented in Section 1. Addressing these gaps will contribute to opti­
specialized knowledge to assess model quality from the perspectives of mizing MHBRs designs to meet the growing demands of urbanization
multiple stakeholders. They also outline a knowledge representation and population growth.
framework, streamlining the maintenance and representation of model-
quality knowledge [63]. Wang et al. propose a Knowledge-Based Design 3. Problem formulation
Guidance System for cloud-based decision support to manage
complexity, uncertainty, and knowledge systematically through a deci­ 3.1. MHBRs’ floor plan design requirements
sion workflow [64]. The contribution is to provide design guidance to
facilitate knowledge discovery, capturing, and reuse in the context of A paramount challenge in automated design generation is ensuring
decision-centric digital design for Industry 4.0 [64]. Nagarajan et al. that outcomes adhere to both functional principles and the pragmatic
propose a novel approach that combines Dimensional Analysis Con­ requirements of architectural design. As delineated in Section 2.1,
ceptual Modeling (DACM) and classical ANNs to develop a new type of architectural design considerations are categorized into three principal
Knowledge-Based ANN (KB-ANN) to address quality control issues in domains: design preferences, design quality, and production re­
Additive Manufacturing (AM) [65]. quirements. Design preferences cater to the customization of architec­
In recent years, generative models, particularly GANs, have emerged tural solutions to align with resident needs and desires. Design quality
as promising alternatives for automated design generation. Hu et al. criteria are centered on the building’s functionality and occupant
[66] proposed Graph2Plan, which uses building boundaries as input to comfort, while production requirements delineate the distinctions be­
output floor plans and their associated layout graphs from the RPLAN tween modular and traditional architectural designs. The design phase
database. Ghannad and Lee [24] developed a coupled generative must address not only the construction phase demands but also the
adversarial network (CoGAN) framework for automated modular production and transportation of prefabricated elements. This study
housing design generation, demonstrating the potential of GANs in this integrates 12 design criteria for MHRBs into the design generation
domain. Nauata et al. [46], and Mohammadreza et al. [67] proposed process, detailed in Table 2. Collectively, these elements constitute a
embedding graph-constrained into the generative adversarial network robust framework for evaluating and designing high-rise residential
that addresses the limitations of existing GAN-based models in handling structures that fulfill both functional and aesthetic considerations and
the unique challenges of high-rise residential building design. Jiang ensure production efficacy.

3
J. Liu et al. Automation in Construction 159 (2024) 105264

Table 1
Comparison of existing methods.
Title Method Output Typical Applications Design knowledge
extraction

House-GAN [46] GAN + Convolutional message Vector Residential+Single flat Yes

passing
Graph2plan: Learning floor plan generation from layout graphs [66] GNN Vector Residential+Single flat Yes
Understanding and visualizing GAN in architectural drawings [50] Pix2Pix Vector Residential+Single flat No
ArchiGAN [69] Pix2Pix Vector Residential+Single flat No
Building layout generation using site-embedded GAN model [49] Pix2Pix Vector Urban+Buildings No
CoGAN [70] Pix2Pix Vector+BIM Modular construction+Single No
flat
Automated layout of modular high-rise residential buildings based Genetic Algorithm Vector MHBRs+One-story No
on genetic algorithm [71]
Automated building layout generation using deep learning and graph U-net Vector+BIM MHBRs+One-story Yes
algorithms [67]
Our proposed method GAN + Convolutional message Vector+BIM MHBRs+One-story Yes
passing

Table 2
MHBRs’ floor plan design knowledge.
Type Requirements Description or Example

Design preference requirements (a) Number-related preferences Desired count of flats, stairs, elevators, and corridors.
(b) Size-related preferences The desired size for each flat.
(c) Existence-related preferences Availability of specific flat type.
Design quality requirements (d) Connectivity All spaces should be interconnected. No isolated space.
(e) Circulation Relationship between two spaces. The path complexity, length, etc.
(f) Compactness Efficient space utilization. Minimize empty or unusable spaces between units.
(g) Brightness Ensure each living space receives ample daylight.
(h) Design norm Example: Standards for windows and doors.
(i) Space quality Each space should have a width-length ratio within a predetermined range.
Design for production requirements (j) Fabrication Ensure maximum dimensions are adhered to, considering material cuts.
(k) Storage Example: Storage space for materials should be 20 cubic meters.
(l) Transportation Ease of moving materials.

3.2. Knowledge representational approaches This method satisfies the design quality requirements by depicting
connections among nodes, as specified in items (d-f) of Table 2. Finally,
This study adopts several approaches to encapsulate design knowl­ the inclusion of dimensional data from actual engineering plans within
edge and requirements. Initially, as illustrated in Fig. 1, a knowledge the generative framework ensures that the spatial quality requirements
graph is employed, wherein each node corresponds to a unique space, and design standards, referenced in items (i) and (h) of Table 2, are
differentiated by colors to indicate space categories, and labeled pragmatic and compatible with real-world production needs, addressing
sequentially based on the quantity of spaces. This model effectively items (j-l) in Table 2.
captures the variety and count of spaces within the floor plan, thereby
fulfilling the representation of design preference requirements outlined
in items (a-c) of Table 2. Subsequently, the design quality requirements 3.3. Evaluation metrics for designs of modular high-rise residential
are articulated through the same knowledge graph. Here, the posi­ buildings
tioning of walls within the graph’s spaces denotes the adjacency re­
lations among them, with interior walls creating interconnected nodes, To validate the effectiveness of architectural floor plans, this study
signifying linkage, while exterior walls suggest no such interconnection. introduces a set of evaluation metrics, as delineated in Table 3. These
metrics are designed to assess both the regularity and rationality of

Fig. 1. Illustration of MHBRs’ design knowledge and related knowledge graph.

4
J. Liu et al. Automation in Construction 159 (2024) 105264

Table 3
Evaluation metrics.
Category Type Requirement Description

Realism(Subjectivity) Connectivity Assesses how seamlessly different spaces within a building are interconnected, ensuring at least one feasible pathway between all pairs of
spaces to facilitate easy movement and interaction.
Circulation Evaluation of the interrelationship between two spatial units in a floor plan, taking into account path length, complexity, and traffic for
each layout path.
Compactness Effective utilization of space to minimize wasted or unusable areas between units, thus maximizing functional space.
Practicality Evaluation of the functional suitability of the layout, including ease of accessibility to key areas and the logical arrangement of utilities and
spaces.
Diversity (Objectivity) None Frechet Inception Distance (FID) score based on rasterized layout images.
Compatibility None Graph edit distance (GED) calculated between the input bubble diagram and the corresponding output layout.
(Objectivity)

partition lines, specifically for stairs, elevators, and flat layouts, as well determine a desired level of diversity. Therefore, this study adopted the
as the overall structural coherence of the entire building layout. Wasserstein GAN (WGAN) model [74] as our baseline for comparison.

• Realism: Realism is assessed by ten seasoned architects, each with a • Compatibility: The compatibility is assessed by computing the Graph
minimum of five years of experience. The assessment utilizes four Edit Distance (GED) [75] between the input bubble diagram and the
principal metrics that uphold structural integrity and functionality: bubble diagram generated from the output layout. The GED score
connectivity, circulation, compactness, and practicality—commonly serves as a standard metric in graph edit distance literature [46,47].
employed in evaluating GAN-generated floor plans as per the refer­ A lower GED score indicates a greater topological similarity between
ences [70, 72]. A 0–10 scoring system quantifies these metrics. the two graph structures.
• Diversity: Diversity is measured using the Frechet Inception Distance
(FID) score, a standard in GAN research [70], following Heusel et al. 4. Methodology
[73]. The FID score serves as a standard metric in GAN literature. It
computes the Wasserstein distance between multidimensional This study presents a sophisticated methodology for the automated
Gaussian distributions of feature vectors derived from real: N(μw − design of floor plans in MHRBs. As illustrated in Fig. 2, the framework
Σ w ) and synthesized images: N(μ − Σ). The FID formula is given by consists of three sequential stages: (1) dataset preparation for floor
Eq. (1). plans, (2) generative modeling for floor plan design, and (3) floor plan
( ) matching followed by BIM generation.
(1)
1
FID = (μ − μw )2 + Tr Σ + Σ w − 2 (ΣΣ w )2
In the first stage, the designer defines constraints such as the number
of stairs, elevators, and types of flats, along with the area of each flat
Lower FID scores signify more remarkable similarity between real
type. Initial processing involves extracting knowledge graphs and floor
and generated samples as per their activation distributions. It is worth
plans from the Gelayout dataset [72], which contains diverse building
noting that there is currently no established threshold FID score to
plans with semantic annotations and graphical structures. Additionally,

Fig. 2. Overview of MHBRs’ floor plan design system.

5
J. Liu et al. Automation in Construction 159 (2024) 105264

Fig. 3. Building floor plan dataset GeLayout with four-channel for presenting the semantic information.

Table 4
Explanation of four-channel labels and values of dataset GeLayout.
Channel Description Value Corresponding Information

Channel 1 Stores the boundary information 127 Boundary

0 Other
Channel 2 Stores the spatial categories 0 Flats
1 Elevators
2 Stairs
3 Corridors
4 Exterior Area
5 Interior Walls
6 Exterior Walls
Channel 3 Distinguishing different rooms with the same labels and number of unique spaces 0 External Areas
1,2, … Room Areas (number of repetitions)
Channel 4 Differentiating the exterior and interior spaces 0 External Areas
255 Internal Areas

the first phase incorporates a floor plan selection method. Building floor 4.1. Floor plan dataset and knowledge graph preparation
plans that align with the designer’s constraints are sourced from the
dataset. This allows the designer to juxtapose them, facilitating a manual 4.1.1. Floor plan dataset and preprocessing
selection based on preference. The study utilizes the GeLayout dataset [72], extracted from DXF files
In the subsequent step, a deep-learning generative model is designed, (data representation format for all information contained in an AutoCAD
trained, and evaluated for the creation of innovative floor plans. The graphics file) containing spatial information, comprising over 700
authors introduce the graph-constrained GAN model called GC-GAN. unique building floor plans. As illustrated in Fig. 3, each floor plan takes
This model leverages the knowledge graph to guide the GAN, enabling the form of a 256 × 256 image in RGBA (Red, Green, Blue, Alpha) format
the production of floor plans that adhere to both the designer’s stipu­ with four distinct channels. These images are not only labeled but also
lations and established design experience. annotated through the four channels, highlighting their semantic in­
In the final stage, pixel-level images are converted into vectorized formation. Channel 1 focuses on boundary delineation. Channel 2 con­
formats and adjusted to meet the designer’s specifications. A flats- veys spatial categories, with varying values corresponding to specific
matching algorithm is then employed in MHRBs, matching the flats spaces, such as flats, elevators, and stairs. Channel 3 stores the codes for
within the standard design library to the core cylinder based on the space categories, with 0 representing the area outside the boundary.
location of each flat’s door. Lastly, through automated BIM generation Spaces of the same type within the boundary are indexed from 1, and the
algorithms, the Revit software serves as a BIM tool, automating the corresponding mask values start from 1. Channel 4 discerns between
creation of 3D building floor plans. This aids architects in their subse­ internal and external areas. The correlations between semantic infor­
quent designs, streamlining the post-processing phases to finalize the mation and their respective values are detailed in Table 4 and Fig. 3.
design. Further details about this methodology can be found in the
subsequent subsections. 4.1.2. Dataset preprocessing
Building on the objective outlined in Section 3.2 to transform
building floor plans into knowledge graphs, the framework initiates with

6
J. Liu et al. Automation in Construction 159 (2024) 105264

Fig. 4. Sample images prepared for training data and corresponding knowledge graphs.

knowledge graph corresponds accurately with the original floor plan,

Table 5 thus meeting design preference requirements. As depicted in Fig. 4,
Color-coding and Channel 2 values for various space types in the original
every corridor links to individual spaces, in instances where the floor
dataset.
plan features two elevators, they remain unconnected to one another.
Space label Channel 2 value in the original dataset R G B A This mirrors the connectivity and circulation for design quality re­
Flat 0 238 77 77 255 quirements. Concurrently, by detecting interior walls and fostering
Elevators 1 198 124 123 255 connections between adjacent spaces, the spatial arrangement of each
Stairs 2 255 210 116 255
area is preserved, ensuring the maintenance of their relative positions
Corridors 3 190 190 190 255
when integrated into the GC-GAN.

a preprocessing phase. Demonstrated in Fig. 4, six exemplar images offer 4.1.3. Floor plan selection with design intention
insight into the training dataset. Initially, the floor plan images undergo As shown in Fig. 5, the study introduces a workflow to identify
a simplification process where space types are reduced to four cate­ compatible floor plans from the dataset based on the design constraints
gories, each distinguished by a specific color as described in Table 5. (proposed floor plan generation system constraints), i.e., the number of
Subsequently, the floor plans are abstracted into knowledge graphs, each space type. The first step of the process is to traverse each sample in
with nodes representing distinct spaces such as stairs, elevators, corri­ the dataset according to the design requirements. The samples that meet
dors, and flats, and edges denoting the requisite connections, contingent the requirements will be provided to the designer for selection. The
on the presence or absence of interior walls. Node classification proceeds second step is manual selection where the designer selects the satisfac­
by interpreting the second channel’s mask value, followed by node tory samples. The process allows multiple selections and flexibility to
number assignment according to the third channel’s mask value. ensure the best input for GC-GAN.
This visualization emphasizes how the number of spaces in the

Fig. 5. The Workflow of the Floor Plan Selected from the Dataset.

7
J. Liu et al. Automation in Construction 159 (2024) 105264

retains a feature in a 1D latent vector space, and a node maintains a

feature volume in Conv-MPN. The convolutional operations in Conv-
MPN serve to update these features within the design space, setting it
apart from typical GCNs. The model details are in the following section.

4.2.1. Generator
The input parameters for the generator algorithm comprise the
segmentation mask of a selected floor plan and its corresponding
knowledge graph. After processing these inputs, the generator produces
an output floor plan, as illustrated in Fig. 7.
Initialization: The first stage of the generator’s operation is
initialization, whereby the knowledge graph and the segmentation mask
of the selected floor plan are used to form the Conv-MPN. In this process,
the authors create a node for each space, utilizing a vector and space
mask to initialize each node, forming a Conv-MPN that mirrors the
structure of the knowledge graph. Each space vector is a 132-d vector →vr ,
→
composed of a 128-d noise vector N concatenating a 4-d space-type
→
vector tr , see in Eq. (2).
{→ }
→
→
vr ← N ; tr (2)

Then, apply a shared linear layer to expand this into a feature volume
of 16 × 8 × 8. The space mask for each node is a conditional image of 2
× 64 × 64. The first channel supplies the segmentation mask, while the
second channel becomes 1 for every pixel when the segmentation mask
is defined. Otherwise, it is 0. Employ a 3-layer CNN to transform the
conditional image into a 16 × 8 × 8 feature volume, which is then
concatenated with the original features to form a 32 × 8 × 8 feature
Fig. 6. Structure of GC-GAN model.
volume that represents each node.
Conv-MPN/Upsampling: The second phase involves executing the
4.2. GC-GAN-based floor plan generation Conv-MPN thrice, coupled with upsampling. The Conv-MPN is respon­
sible for refining the feature volume pertaining to space through a
To facilitate the generation of novel floor plans based on the design method known as convolutional message passing. This updating tech­
preferences outlined in Section 4.1.3, this research proposed a network nique is detailed as follows: Initially, features of spaces that are inter­
structure called GC-GAN. Drawing inspiration from House-GAN++ connected within the graph are concatenated. Following this, the
[47], the researchers introduced the GC-GAN concept. As depicted in features of spaces that are not connected are concatenated. Lastly, this
Fig. 6 and influenced by House-GAN++, GC-GAN operates as a rela­ amalgamated data undergoes processing via CNN (Eq. (3)).
tional generative adversarial network featuring both a generator and a [ ]
discriminator. The role of the generator is to produce authentic-seeming vlr ←CNN vlr ; Poolvls ; Pool vls (3)
synthetic data that aids in the training of the discriminator. At the same s∈c(r) →
s∈ c (r)
time, the discriminator is trained to differentiate between the synthetic
outputs of the generator and genuine data. The discriminator metes out Where [;,;] indicates concatenation, vlr feature volumes for a
penalties if the generator’s results appear inauthentic. Finally, through component r. l indicates the number of times the operation Conv-MPN is
backpropagation, the decisions of the discriminator serve as feedback performed. c(r) and → c (r) denote sets of spaces that are connected and
for the generator. The generator can leverage this signal to optimize its not connected, respectively. Concurrently, during the upsampling stage,
weights, thus enhancing the data it generates. the dimensions of the features are augmented by a factor of 2 through
The network backbone is a convolutional message-passing network the utilization of a transposed convolution with a kernel size of 4, a
(Conv-MPN [76]). Conv-MPN can be seen as a unique variant of Graph stride of 2, and padding of 1. This is done while maintaining the same
Convolutional Networks (GCNs) [77]. GCNs are a class of neural net­ number of channels. Consequently, the final feature volume is size 16 ×
works designed specifically for processing data structured as graphs, 64 × 64.
capturing relationships and features of nodes and edges effectively. The Convolution: The final phase for the generator is convolution. A
main distinction between Conv-MPN and GCNs lies in the feature rep­ shared three-layer CNN is used to convert the final feature volume into a
resentation and updating mechanism. Unlike traditional GCNs, a node space segmentation mask of size 1 × 64 × 64, thereby outputting the

Fig. 7. Structure of GC-GAN generator.

8
J. Liu et al. Automation in Construction 159 (2024) 105264

Fig. 8. Structure of GC-GAN discriminator.

new floor plan.

Table 6
GC-GAN architectural specification.
4.2.2. Discriminator
The input of the discriminator is a floor plan segmentation mask, Architectural Layer Specification Output Size
which comes from the generator or ground-truth samples, and the Generator Conacat 1 128 + 4 1 × 132
output is a scalar Score, classifying ground-truth samples from gener­ Linear reshape 132 × 1024 8 × 8 × 16
ated ones. The discriminator performs a series of operations in the same Conacat 2 8 × 8 × 16 + 8 × 8 × 16 8 × 8 × 32
Conv-MPN 1 Stride = 2 Padding = 1 8 × 8 × 16
order, as shown in Fig. 8. Up sampling 1 Scale factor = 2 16 × 16 × 16
Initialization: The Conv-MPN is formulated by the discriminator Conv-MPN 2 Stride = 2 Padding = 1 16 × 16 × 16
through the utilization of a floor plan segmentation mask and a space- Up sampling 2 Scale factor = 2 32 × 32 × 16
→ Conv-MPN 3 Stride = 2 Padding = 1 32 × 32 × 16
type vector tr . The segmentation mask possesses a size of 1 × 64 ×
Up sampling 3 Scale factor = 2 64 × 64 × 16
64. In order to expand a 4-D space-type vector, apply a linear layer to Conv-MPN 4 Stride = 2 Padding = 1 64 × 64 × 16
expand it to 32,768-d. Subsequently, the vector is reshaped into an 8 × Discriminator Linear reshape 4096 × 32,768 64 × 64 × 8
64 × 64 tensor. This tensor is concatenated with the segmentation mask, Conacat 1 64 × 64 × 8 + 64 × 64 × 1 64 × 64 × 9
forming a feature volume of 9 × 64 × 64 to symbolize each node. Conv-MPN 1 Stride = 2 Padding = 1 64 × 64 × 16
Down sampling 1 Scale factor = 0.5 64 × 64 × 16
Conv-MPN/Downsampling: The discriminator Conv-MPN process Conv-MPN 2 Stride = 2 Padding = 1 32 × 32 × 16
is the same as generator. During the downsampling phase, the feature Down sampling 2 Scale factor = 0.5 32 × 32 × 16
dimensions are effectively reduced by a factor of 2 through the imple­ Conv-MPN 3 Stride = 2 Padding = 1 16 × 16 × 16
mentation of a standard convolution process with a kernel size of 3, a Down sampling 3 Scale factor = 0.5 8 × 8 × 16
Conv-MPN 4 Stride = 2 Padding = 1 8 × 8 × 16
stride of 2, and padding of 1. This reduction is carried out while ensuring
the number of channels remains constant. As a result, the ensuing
feature volume after this process will have dimensions of 16 × 8 × 8. D(̃x) is the discriminator’s prediction of the sample generated by the
Convolution: A duo of convolution operations is carried out, one generator G. The goal of the value function is to minimize the score of
preceding and one following the three rounds of Conv-MPN and the the generated sample and maximize the score of the real sample. The
downsampling process. During the initial operation, the feature volume loss functions of the generator and the discriminator are given by Eq. (6)
is transformed into a size of 16 × 32 × 32 through a shared three-layer and Eq. (7).
CNN. Subsequently, another three-layer CNN is employed to alter the
̅→ 1∑
final feature volume into a 128-d vector dr . x) ] + λ
L(G) = − E [D(̃ |mi − m
̂ i| (6)
x̃∼Pg N iϵN
( )
̅→
Score←Linear ​ Pool dr ​ (4) When the segmentation mask is specified, perform an L1-loss be­
r
tween the conditional image mi and the generated mask m ̂ i . i represents
In the final step, aggregate all the space vectors using sum-pooling the index of a node, and N represents the set of all nodes. λ is a constant
and introduce a single linear layer to produce a scalar score as shown set to 1000.
in Eq. (4). This scalar score distinguishes between the ground-truth [( )2 ]
samples and the ones generated by the model. L(D) = E [D(̃x) ] − E [D(x) ] + λ E ‖∇x̂ D(̂ x ) ‖2 − 1 (7)
x̃∼Pg x∼Pr x̂∼Px̂

4.2.3. GC-GAN training Px̂ is a distribution obtained by uniformly sampling along the line
Our model employs Wasserstein GANs-Gradient Penalty (WGAN-GP) between the real distribution Pr and the generated distribution Pg. λ is
loss [74] with a gradient penalty coefficient set to 10. The gradient the gradient penalty coefficient, set to 10, followed by Gulrajani et al.
( )2
penalty term was implemented through linear, uniform interpolation [74]. ‖∇x̂ D( ̂x ) ‖2 − 1 represents the gradient penalty term.
between real and generated sample segmentation mask pixels. The The experiments are conducted using Python 3.9 and PyTorch 1.13
following formulas will elaborate on the training process of GC-GAN. on an NVIDIA GeForce RTX 3080 with 10 GB of memory. The detailed
The value function of the model is defined in Eq. (5). configuration of the model layers is shown in Table 6. The model used
V (GD) = minmax E [D(x) ] − E [D(̃
x) ] (5) ADAM as the optimizer and was trained for 500 iterations. The learning
G D x∼Pr x̃∼Pg rates for both the generator and the discriminator were set at 0.0001,
Where E represents expectation, G and D, respectively, represent the with the first and second momentum parameters defined as 0.5 and
generator and the discriminator. Pr represents the real data distribution, 0.999, respectively. The batch size is 1. Our training strategy involved
and Pg represents the distribution generated by the model defined by ̃x = setting the number of critics to 1, meaning that for every training cycle,
G(z). According to the definition, z represents a noise vector standard updating the discriminator and the generator once each to maintain
normal distribution, and x represents a sample from the real data dis­ equilibrium. For all non-linearities, leaky-ReLUs(α = 0.1) were
tribution. D(x) is the discriminator’s prediction of the real sample and employed, except for the final layer in the generator where a hyperbolic

9
J. Liu et al. Automation in Construction 159 (2024) 105264

Fig. 9. Standard 3D model design library building process.

Fig. 10. Standard 3D model design library.

10
J. Liu et al. Automation in Construction 159 (2024) 105264

tangent function was used. based on the information derived from the vector drawings, as illus­
trated in Fig. 9. The parameterization process, as delineated in Fun’s
method [78], involves extracting the outer frame of the flat, as well as
4.3. Floor plan matching with standard design library and BIM generation
determining the positions of doors, windows, and lighting fixtures. This
process serves to simplify the irregular floor plan into a regular, vec­
The images generated by the GC-GAN can exhibit irregular shapes.
torized rectangle by utilizing four boundary axes as the edges of the
To address this issue and ensure the practical suitability of these images
rectangles. Moreover, the locations of the doors and windows are also
for modular design, this study introduces a matching algorithm. The
mapped to the edges of the resulting rectangle. The standard design li­
floor plans generated by GC-GAN are parametrically transformed into
brary of floor plans includes different areas to accommodate a range of
vector diagrams based on the data from the actual project and matched
requirements, as shown in Fig. 10. It contains three types of areas: 65 m2,
with the modular flats from the standard design library. The matching
80 m2, and 100 m2. For the 80 m2 and 100 m2 categories, there are 3
outcome is contingent on design preferences, taking into consideration
modular flat designs, while for the 65 m2 category, there are two designs
various design requirements such as connectivity, brightness, and pro­
available. These diverse design options are tailored to meet distinct
duction criteria during the matching process. This approach will serve as
living needs.
an effective blueprint for designers in subsequent design phases. Ulti­
mately, an automated BIM generation algorithm will convert the floor
4.3.2. Image vectorization and matching algorithm
plan into a 3D model in Revit. Further elaboration is provided in the
This paper presents an image vectorization and matching algorithm
subsequent sections.
developed to convert pixel-based floor plans generated by GC-GAN into
vectorized designs suitable for architectural applications. As depicted in
4.3.1. Standard design library with modularity
Fig. 11, the methodology consists of two main stages. The initial stage,
To address design requirements related to production and spatial
termed the “vectorization core,” comprises four sequential steps: seg­
quality, this study has developed a standardized design library. The
mentation, boundary identification, modulation, and collision avoid­
initial step involves parameterizing the CAD drawings of modular units
ance. In contrast, the second stage, referred to as “match flats,”
to generate vector drawings. Subsequently, BIM models are constructed

Fig. 11. Image vectorization and flat-matching algorithm process.

11
J. Liu et al. Automation in Construction 159 (2024) 105264

encompasses boundary identification for flats, modularization, flat-core proceeds to place the appropriate BIM object at each unit’s center,
tube matching, and collision avoidance. applying rotation to flat units where necessary.
In the first stage, image segmentation is performed by identifying the
Algorithm 1. Automated BIM generation algorithm.
colors in different spaces to extract the core domain. The bounding box

for each space and core region is then determined using the OpenCV 5. Experimental results and discussions
library [79]. In the modalization process, transforms pixels to vectors,
adjusting the number of pixels in the center of each space to match its This paper employs five scenarios, guided by design preference re­
actual size and a predefined aspect ratio. This ratio is established from quirements, to illustrate the efficacy of the proposed automated floor
CAD drawings in the standard design library. For example, the stair plan design framework for MHBRs. As shown in Fig. 12, the number of
space is designated as 21 m2 with a 3:1 aspect ratio, while the elevator units for scenarios A-E ranges from two to six depending on the con­
space is 4 m2 with a 1:1 ratio. Collision avoidance is the final step in this straints (stairs, elevators, and number of units) required by the designer.
stage, ensuring there are no spatial overlaps. Three different floor plans were selected for each scenario. Within the
The second stage commences with the identification of boundaries same case, there were variations in spatial relations or quantity differ­
for each flat, capturing specific parameters such as bounding boxes and ences. Case-1 and Case-2 have the same or similar knowledge graphs,
centroid coordinates. Flats are then subjected to modularization based and Case-3 has a different knowledge graph, see Fig. 14 ‘Real’ column
on their dimensions as outlined in the standard design library, the and ‘Knowledge graph’ column.
process same to the initial stage. Subsequently, the flats are aligned with Based on the provided cases, Section 5.1 introduces the results of
the midpoint of the core tube’s edge corresponding to their spatial ori­ floor plan generation and evaluates the generated results based on the
entation—top, bottom, left, or right. The final step involves collision proposed evaluation measures. Section 5.2 presents the outcome of the
avoidance, distinct from the first stage, where corridors that can be floor plan matching the standard design library and automated BIM
linked to flats are determined by door locations, ensuring that stairs or generation. Lastly, Section 5.3 offers a discussion of the results and
elevators do not interfere within an eight-pixel range. This measure findings. These sections collectively represent the comprehensive vali­
maintains connectivity between flats and corridors. In cases where dation process applied to assess the effectiveness of the proposed auto­
overlaps are detected during the collision check, the algorithm re­ mated design framework.
positions the flat along the core tube’s edge to resolve the issue. This
matching protocol ensures that doors maintain access to corridors and
5.1. GC-GAN-based floor plan generation
that window lighting is preserved. In cases where two flats compete for
the same edge position, the algorithm selects the 65 m2 flat by default.
In evaluating the performance of the GC-GAN model, three metrics
This decision is based on the standard design library’s specification that are utilized: realism, diversity, and compatibility. More detailed, the
the 65 m2 flat has windows on only one side, unlike other models with
generated floor plan’s realism consists of connectivity, circulation,
windows on both sides. compactness, and practicality. Diversity and compatibility are quanti­
This algorithm aims to assist designers during the initial design
fied using the FID score and GED score, respectively. A survey was
phases, simplifying the core tube’s shape into a rectangular represen­ designed, employing a rating scale ranging from 0 (Poor) to 10 (Excel­
tation through the generation process. In the post-processing stage, de­
lent). This survey sought feedback from experienced designers, each
signers can flexibly refine the design outcome by adjusting the position possessing a minimum of five years of expertise in the field of MHRBs
or shape of various spaces in Revit to achieve the final design, as elab­
design. A total of ten seasoned designers were invited to assign scores to
orated in Section 5.2.2 for detailed information. each realism indicator.

4.3.3. Floor plan BIM generation 5.1.1. Evaluation metrics of realism for testing scenarios
This study introduced an automated BIM generation algorithm,
The scores for the four indicators of realism under five scenarios can
developed in Python and utilizing the Revit Application Programming be seen in Fig. 13(a). Table 7 displays the scores for five different model-
Interface (API) 2023. Presented as pseudocode in Algorithm 1, the
generated outcomes, with each value representing the average score
program’s operational framework is detailed. It processes two inputs: a from ten participants. Generally, the average scores for all cases surpass
parameterized floor plan and a modular flats BIM library. The algorithm
6, indicating that most model-generated results meet design re­
categorizes the given space into core area units(including stairs, eleva­ quirements. The lowest-scoring case is 5.88 (with a variance of 0.36),
tors, and corridors) and flat units. For core area units, it retains the x and
and the highest-scoring case is 6.93 (with a variance of 0.47), with the
y coordinates of the central point and the designated type. For flat units, overall variance being 0.37. The stable scores between cases and a
it holds the central coordinates and the rotation angle. The algorithm
minimum score near 6 suggest that the model generation effects are

12
J. Liu et al. Automation in Construction 159 (2024) 105264

Fig. 12. All samples’ floor plans and knowledge graphs.

steady, but there is room for improvement. Among the evaluation in­ 5.1.2. Evaluation metrics of diversity and compatibility for testing scenarios
dicators, the total scores for Connectivity and Circulation are 6.75 and Regarding diversity and compatibility, the authors assessed 400
6.7, respectively, while the scores for Compactness and Practicality are randomly chosen pairs of generated and authentic images, employing
6.21 and 6.11, being lower by 0.5. This reveals that designers are FID and GED to gauge these attributes. Fig. 13(b) displays the mean
satisfied with the relative positions of spaces but are relatively unsatis­ scores for realism, diversity, and compatibility: an FID score of 27.41
fied with the regularity and gaps of layout and spaces. The overall scores indicates a high level of image quality, signifying a closer resemblance to
not reaching the satisfactory level are mainly due to the irregular shape real images compared to the House-GAN++ model’s average FID score
of spaces and gaps, but the satisfaction of relative positional relation­ of 32.05. Generally, a lower FID score is better, as it denotes smaller
ships guarantees the rationality of the subsequent floor plan matching differences between the sets of images being compared. Similarly, a
with the standard design library process. While maintaining the relative lower GED score is desirable, representing a greater topological simi­
positional relationships of each space, spaces are regularized and vec­ larity between graphs. In this case, the GED score for our model is 11.99,
torized based on actual modular design requirements. which, unlike the FID score, does not improve upon the House-GAN++
model’s average GED score of 2.6. This is attributed to our model’s more

Fig. 13. Comparison of the proposed metrics of five test scenarios.

13
J. Liu et al. Automation in Construction 159 (2024) 105264

Table 7
Evaluation metrics of realism for five testing scenarios (connectivity, circulation, compactness and practicality).
Scenarios Case Metrics SUM AVG SD

Connectivity Circulation Compactness Practicality

A 1 7.2 7.3 6.5 6.5 27.50 6.88 0.38

2 7.1 7.6 6.6 6.4 27.70 6.93 0.47
3 7 7 6.8 5.9 26.70 6.68 0.45
B 1 6.9 7.3 6.3 6.1 26.60 6.65 0.48
2 7 7.4 6.6 6.1 27.10 6.78 0.48
3 6.6 6.9 6.3 6.1 25.90 6.48 0.30
C 1 7.6 7.3 6.1 6.9 27.90 6.98 0.56
2 6.9 6.4 5.7 6.2 25.20 6.30 0.43
3 6.2 6 5.9 6.1 24.20 6.05 0.11
D 1 6.4 6 5.5 5.6 23.50 5.88 0.36
2 6.4 6.1 5.8 5.5 23.80 5.95 0.34
3 6.3 6 5.3 6 23.60 5.90 0.37
E 1 6.7 6.2 6.7 6.2 25.80 6.45 0.25
2 6.9 6.8 6.5 6.5 26.70 6.68 0.18
3 6.1 6.2 6.6 5.6 24.50 6.13 0.36
SUM 101.30 100.50 93.20 91.70 \ 96.68 5.51
AVG 6.75 6.70 6.21 6.11 25.78 6.45 0.37
SD 0.40 0.57 0.45 0.36 1.50 0.37

intricate graph structure, which incorporates additional nodes and choose the desired flat type. The requirements for design specifications,
connections. Our dataset comprised floor plans, unlike the flats in the spatial quality, and compactness are addressed while maintaining the
House-GAN++ dataset. Fig. 14 presents the learning outcomes of the design quality requirements for connectivity and circulation of GC-GAN-
GC-GAN, offering qualitative evidence of compatibility. In calculating generated results. Meanwhile, the use of modularized standard flats
GED, the number of unsatisfied graph constraints was also tallied, with a model ensures the generated designs meet the production and transport
result of zero, underscoring the model’s reliability in adhering to input requirements.
constraints. Fig. 15 shows the results of matching multiple standardized flat
types. Using scenarios, A and B as examples, the three examples in
5.1.3. Comparison of the generated and real floor layouts scenario A complete the matching with flat-80 A, flat-80B, and flat-80C,
An analysis of scenarios A through E depicted in Fig. 14 reveals that, respectively. All cases in Scenario B complete the matching with flat-
with the exception of scenario D, the generated images consistently 80C. This demonstrates that the matching algorithm can fully satisfy
maintain the relative spatial relationships observed in the original floor the diversity of choices while ensuring stability under the same choice.
plans. In scenario D, the elevator and staircase are significantly shifted, Parallel rooms appear in scenarios C, D, and E, and according to the
but this alteration does not affect the consistency of the knowledge algorithm library case, to ensure the lightness of the windows, flat-65 A
graph of the generated image with the original floor plan, which in­ is the choice for the side-by-side rooms. Under this premise, flat-100 A,
dicates that the GC-GAN model is able to adapt to a variety of design flat-100B, and flat-100C matches were used sequentially for scenario C,
constraints, such as the number of spaces, connectivity, and layout. A and all flat-100B matches were used for scenario D. This demonstrates
granular examination of Case-1 and Case-2 across these scenarios in­ the versatility and stability of the algorithm in complex situations.
dicates that the GC-GAN yields stable outcomes across differing
authentic floor plans, suggesting that the spatial configuration exerts 5.2.2. BIM generation and post-processing output
minimal impact on the results. Moreover, when comparing Cases 1 and 2 Based on the results from Section 5.2.1, vector information is
to Case 3 within the same scenarios, it is evident that the GC-GAN is extracted from the images for generating BIM models. The generated
capable of producing distinct designs from similar floor plans in vectorized images can be accurately transformed into BIM models, as
response to differing knowledge graphs, thereby demonstrating its demonstrated in the BIM generation example (E-case3) shown in Fig. 16.
flexibility to suit varied design needs. This process ensures a precise transition from vectorized images to en­
gineering software. Post-processing is necessary due to the difficulty of
complying with all design specifications. Leveraging the highly inte­
5.2. Floor plan matching with standard design library and BIM generation grated platform of Revit, engineers can directly perform simple post-
processing tasks, such as repositioning various components and adjust­
5.2.1. Floor plan matching with standard design library ing core shapes and plane positions to meet their unique requirements.
Following the algorithmic operations detailed in Section 4.3.2, the As shown in the post-processing example in Fig. 17, the designer can
spatial information is extracted from the image and transformed into a visually adjust the position of the elevator module and stair module to
vectorized figure that meets the design requirements. The model is complete the design. The entire process is straightforward and efficient,
completed in about 5 s of generating a single design result. Fig. 15 bypassing the traditional modeling steps and enabling engineers to
demonstrates the effectiveness of the floor plan matching process, with modify the visualized models directly, ultimately achieving the output
the ‘Generate’ row showing the results generated by GC-GAN and the of architectural blueprints. The initial BIM model serves as a guide to
‘Match’ row showing the results generated by the matching algorithm. It ensure vector accuracy and simplify user parallel displacement adjust­
is evident that the post-matched vector image maintains the relative ments. This design process transformation into an intuitive operation
positions of each space as in the original GC-GAN-generated images. As accelerates the design workflow, enhances efficiency, and reduces
explained in Section 4.3.2, the vector image takes into account the errors.
design requirements for door and window placements, ensuring smooth Fig. 18 (a) and (c) showcase the plan view generation results of the
circulation between the core area and individual units, as well as proper entire design system, along with the example of its 3D models. The
lighting within each unit. Each pixel in the image represents the actual image-matching process utilizes the standard design library, ensuring
area. Moreover, the matching process considers the preferences of the that the generated floor designs comply with modular building
designer. While ensuring adequate lighting conditions, the designer can

14
J. Liu et al. Automation in Construction 159 (2024) 105264

requirements, guaranteeing the usability of the design output. Fig. 18 “IfcMaterialLayer” indicates not only the material used but also its
(b) displays part of the model’s Industry Foundation Classes (IFC) file, thickness, which directly translates to production material re­
showing partial information about the walls. The model’s IFC file shows quirements. Similarly, “IfcMaterialLayerSetUsage” defines the posi­
wall properties, thickness, length, and other information. Compliance tioning and orientation of these wall material layers, offering a blueprint
with the modularization is demonstrated, and production requirements for assembly sequences. Additional attributes such as “LoadBearing” and
are addressed. The potential of IFC extends beyond mere data interop­ “IsExternal,” obtained from entities like “IfcPropertySingleValue” and
erability in the AEC industry. It offers critical insights that can guide the “IfcPropertySet,” further inform the manufacturing process by speci­
production of prefabricated building components. The IFC schema in­ fying the structural and environmental requirements of each wall
cludes key entities like “IfcMaterialLayer”, “IfcMaterialLayerSetUsage”, component. By mapping these IFC entities to the corresponding modules
and “IfcWallType” which hold valuable information about the material in the assembly line, a more streamlined, efficient, and error-free pro­
properties, alignments, and design types of walls. For instance, the entity duction process can be achieved. Thus, IFC serves as an instrumental

Fig. 14. The generation results of GC-GAN.

15
J. Liu et al. Automation in Construction 159 (2024) 105264

Fig. 14. (continued).

tool for bridging the gap between digital design and physical production construction’s growing significance, especially in the early design
in the realm of modular construction. stages, integrating design and construction processes is imperative. The
proposed framework in this paper facilitates the automatic trans­
5.3. Discussion formation of floor plan designs into three-dimensional models through
the integration of floor plan design and BIM post-processing. This en­
This research introduces an innovative floor plan design framework ables designers to fully visualize and follow the design-to-production
for MHRBs, surpassing conventional design techniques. With modular continuum. The framework presents an all-encompassing strategy,

16
J. Liu et al. Automation in Construction 159 (2024) 105264

Fig. 15. The matching results of the floor plan.

incorporating production, manufacturing, and transportation consider­ process. In contrast, the proposed framework can be completed in just
ations to ensure that floor plans are tailored to the distinctive re­ 20 min. Moreover, unlike the homogenized tendency of traditional
quirements of modular construction. The proposed data-driven design, the data-driven framework can generate diverse results using
framework significantly outperforms traditional manual design, which GC-GAN. Meanwhile, designers can choose their preferred flat types
takes nearly 4 days to complete the floor plan design and modeling from the standard library to enhance the diversity and personalization of

17
J. Liu et al. Automation in Construction 159 (2024) 105264

Fig. 16. An Instance of BIM generation.

Fig. 17. Illustrations of user post-processing.

the design, while maintaining the quality of the outcomes. 6. Conclusion and future work
Nevertheless, the current design constraints addressed are founda­
tional, excluding aesthetic preferences and catering only to elementary Efficient and rapid architectural residential floor plan design has
design situations. There is potential for enriching the diversity and always been a pivotal concern in the AEC industry. While MHBRs, owing
thoroughness of design constraints to cover a broader spectrum of design to their distinct advantages, have been extensively embraced, the design
scenarios. BIM model generation promotes consistent information flow process is not devoid of challenges. Key issues include the need for
throughout the project’s lifespan, which is vital for modular construc­ design and production coordination, a perceived lack of design flexi­
tion’s need for uninterrupted data exchange from design to build. In this bility, and a heavy reliance on the designer’s experience to ensure
light, our framework acts as a robust facilitator of this process. None­ quality outcomes. Given these challenges, this study utilizes the capa­
theless, integration is constrained by the capabilities of the Revit API, bilities of GC-GAN to propose an automatic building floor plan design
restricting our post-processing to this specific architectural software. framework for MHBRs. The innovative aspects of this framework
As MHRBs exemplify the shift toward modularization in residential contribute to knowledge and practice in several ways:
construction, our framework streamlines the transition from traditional
to modular residential formats. It serves as a model for the modulari­ 1. This research undertook a meticulous analysis of the MHBRs’ floor
zation of various building types, offering a valuable reference for diverse plan to extract critical design knowledge and determine relevant
modular construction projects. A key attribute of our framework is its evaluative metrics, establishing a foundation for intelligent design of
deference to designers’ preferences, recognizing that generative design floor plans, including design preference requirements (layout num­
transcends automation and efficiency to embrace design quality and ber, size, and existence), design quality requirements (connectivity,
variety. While the assortment of design options is currently limited by circulation, compactness, brightness, design norm, and space qual­
the selection in the standard design library, the scope for augmenting ity), and also production requirements (fabrication, storage, and
design diversity expands as these collections grow. transportation).
2. This research developed a graph structure representation to encap­
sulate the intricate design knowledge of MHRBs. This approach

18
J. Liu et al. Automation in Construction 159 (2024) 105264

Fig. 18. An instance of final design outputs and IFC information.

utilizes a graph structure to render complex design concepts into 4. The study also proposes a matching method based on a standard
interpretable and operational knowledge. In this graph, nodes sym­ design library. The method can vectorize the design results and
bolize individual spaces, while edges define the connections between match them with the standard designs in the design library. It solves
them. This method effectively conveys design preferences and the randomness of the results generated by the traditional GAN and
quality requirements by delineating various space types in the floor ensures that the design results meet the design specifications and
plan, their respective quantities, and the circulation dynamics among actual production requirements. At the same time, this method
them. provides personalized designs based on the designer’s preferences.
3. To address the generation of a new floor plan based on previous As long as the design library is rich in styles, endless design results
design knowledge, the research presents a framework with a GC- can be created. The research proposed that the BIM auto-generation
GAN model based on graph constraints for intelligently designing method can save a substantial amount of modeling time and visualize
floor plans for MHBRs. Using this framework, designers can generate the design process. Using the highly integrated Revit software, de­
multiple floor plans after setting the number of elevators, flats, and signers can conveniently adjust the design according to their needs
stairs. The method uses various MHBRs’ design rules as constraints to and output the design results to IFC components to ensure that the
generate designs that satisfy the design requirements and facilitate design results meet the production and transportation requirements
post-processing by the designer. Through training, the GC-GAN of MHBRs.
model generates new floor plans by the graph structure represent­
ing the design requirements as constraints. Experimental results Though the focal point of this research is MHBRs, the data-driven
show that the model can quickly generate diverse flat maps that meet generative design framework accentuates spatial design considerations
the design requirements. in the early design phase, offering relevance across a spectrum of

19
J. Liu et al. Automation in Construction 159 (2024) 105264

housing projects, irrespective of their construction methodologies. This [10] V.J.L. Gan, BIM-based graph data model for automatic generative design of
modular buildings, Autom. Constr. 134 (2022) 104062, https://doi.org/10.1016/j.
universality signifies the framework’s potential in modular architecture
autcon.2021.104062.
and even traditional residential and commercial building projects, [11] Z. Zheng, Y. Li, J. Torres, An integrated method of automated layout design and
promising a revolution in design efficiency and automation. The broader optimization for modular construction, engineering, Construct. Architect. Manage.
applicability of the proposed methodology hints at its potential incor­ (2022), https://doi.org/10.1108/ECAM-04-2022-0329.
[12] Z. Wang, W. Pan, Z. Zhang, High-rise modular buildings with innovative precast
poration into various project types. concrete shear walls as a lateral force resisting system, Structures. 26 (2020)
Future research directions include refining the GC-GAN model to 39–53, https://doi.org/10.1016/j.istruc.2020.04.006.
better incorporate additional design constraints, such as sustainability [13] A.W. Lacey, W. Chen, H. Hao, K. Bi, Review of bolted inter-module connections in
modular steel buildings, J. Build. Eng. 23 (2019) 207–219, https://doi.org/
and energy efficiency. Another possible direction is to explore inte­ 10.1016/j.jobe.2019.01.035.
grating the proposed framework with BIM systems to streamline the [14] R.M. Lawson, R.G. Ogden, R. Bergin, Application of modular construction in high-
design and construction processes further. rise buildings, J. Archit. Eng. 18 (2012) 148–154, https://doi.org/10.1061/(ASCE)
AE.1943-5568.0000057.
[15] Y.-F. Liu, S.-S. Luo, H. Wang, Research on a complete set of technologies for
CRediT authorship contribution statement assembled residential buildings with steel-structure based on house type
modularization and component standardization, in: The 7th International
Conference on Environmental Science and Civil Engineering, Series: Earth and
Jiepeng Liu: Writing – review & editing, Project administration, Environmental Science, 2021, pp. 719–022063, https://doi.org/10.1088/1755-
Funding acquisition. Zijin Qiu: Writing – review & editing, Writing – 1315/719/2/022063.
original draft. Lufeng Wang: Formal analysis, Data curation, Concep­ [16] W. Ferdous, Y. Bai, T.D. Ngo, A. Manalo, P. Mendis, New advancements, challenges
and opportunities of multi-storey modular buildings – a state-of-the-art review,
tualization. Pengkun Liu: Writing – review & editing, Writing – original Eng. Struct. 183 (2019) 883–893, https://doi.org/10.1016/j.
draft, Visualization, Supervision, Methodology. Guozhong Cheng: Re­ engstruct.2019.01.061.
sources, Project administration. Yan Chen: Investigation. [17] S. Wu, N. Zhang, X. Luo, W.-Z. Lu, Intelligent optimal design of floor tiles: a goal-
oriented approach based on BIM and parametric design platform, J. Clean. Prod.
299 (2021) 126754, https://doi.org/10.1016/j.jclepro.2021.126754.
[18] S. Wu, N. Zhang, X. Luo, W.-Z. Lu, Multi-objective optimization in floor tile
Declaration of competing interest planning: coupling BIM and parametric design, Autom. Constr. 140 (2022)
104384, https://doi.org/10.1016/j.autcon.2022.104384.
[19] J. Lee, M. Park, H. Lee, M. Kim, H. Hyun, Development of conceptual modular
The authors declare that they have no known competing financial
building unit design framework for inexperienced designers at the pre-design stage,
interests or personal relationships that could have appeared to influence modular and offsite construction (MOC), Summ. Proc. (2015) 188–195, https://
the work reported in this paper. doi.org/10.29173/mocs151.
[20] H. Liu, Y.-L. Yang, S. AlHalawani, N.J. Mitra, Constraint-aware interior layout
exploration for pre-cast concrete-based buildings, Vis. Comput. 29 (2013)
Data availability 663–673, https://doi.org/10.1007/s00371-013-0825-1.
[21] S. Feuerstack, M. Blumendorf, V. Schwartze, S. Albayrak, Model-based layout
Data will be made available on request. generation, in: Proceedings of the Working Conference on Advanced Visual
Interfaces, Association for Computing Machinery, New York, NY, USA, 2008,
pp. 217–224, https://doi.org/10.1145/1385569.1385605.
Acknowledgments [22] M. Inoue, H. Takagi, Layout algorithm for an EC-based room layout planning
support system, in: 2008 IEEE Conference on Soft Computing in Industrial
Applications, 2008, pp. 165–170, https://doi.org/10.1109/SMCIA.2008.5045954.
The authors would like to express their gratitude for the financial [23] K. Grzesiak-Kopeć, B. Strug, G. Ślusarczyk, Evolutionary methods in house floor
support provided by the Technology Innovation and Application plan design, Appl. Sci. 11 (2021) 8229, https://doi.org/10.3390/app11178229.
[24] P. Ghannad, Y.-C. Lee, Automated modular housing design using a module
Development of Chongqing Science and Technology Commission
configuration algorithm and a coupled generative adversarial network (CoGAN),
(CSTB2022TIAD-KPX0136) and the National Natural Science Founda­ Autom. Constr. 139 (2022) 104234, https://doi.org/10.1016/j.
tion of China (No. 52130801). The opinions expressed in this paper are autcon.2022.104234.
solely those of the authors. [25] X. Kong, L. Jiang, H. Chang, H. Zhang, Y. Hao, H. Gong, I. Essa, BLT: Bidirectional
layout transformer for controllable layout generation, in: S. Avidan, G. Brostow,
M. Cissé, G.M. Farinella, T. Hassner (Eds.), Computer Vision – ECCV 2022, Springer
References Nature Switzerland, Cham, 2022, pp. 474–490, https://doi.org/10.1007/978-3-
031-19790-1_29.
[26] K. Wang, Y.-A. Lin, B. Weissmann, M. Savva, A.X. Chang, D. Ritchie, PlanIT:
[1] H.-T. Thai, T. Ngo, B. Uy, A review on modular construction for high-rise buildings,
planning and instantiating indoor scenes with relation graph and spatial prior
Structures. 28 (2020) 1265–1290, https://doi.org/10.1016/j.istruc.2020.09.070.
networks, ACM Trans. Graph. 38 (2019), https://doi.org/10.1145/
[2] W. Pan, Y. Yang, L. Yang, High-rise modular building: Ten-year journey and future
3306346.3322941, pp. 132:1–132:15.
development, in: Construction Research Congress 2018, American Society of Civil
[27] K. Wang, M. Savva, A.X. Chang, D. Ritchie, Deep convolutional priors for indoor
Engineers, New Orleans, Louisiana, 2018, pp. 523–532, https://doi.org/10.1061/
scene synthesis, ACM Trans. Graph. 37 (2018), https://doi.org/10.1145/
9780784481301.052.
3197517.3201362, pp. 70:1–70:14.
[3] H. Liu, C. Sydora, M.S. Altaf, S. Han, M. Al-Hussein, Towards sustainable
[28] J. Li, J. Yang, A. Hertzmann, J. Zhang, T. Xu, LayoutGAN: Generating Graphic
construction: BIM-enabled design and planning of roof sheathing installation for
Layouts with Wireframe Discriminators, 2019, https://doi.org/10.48550/
prefabricated buildings, J. Clean. Prod. 235 (2019) 1189–1201, https://doi.org/
arXiv.1901.06767.
10.1016/j.jclepro.2019.07.055.
[29] J. Li, J. Yang, J. Zhang, C. Liu, C. Wang, T. Xu, Attribute-conditioned layout GAN
[4] Y. Song, R. Koeck, S. Luo, Review and analysis of augmented reality (AR) literature
for automatic graphic design, IEEE Trans. Vis. Comput. Graph. 27 (2021)
for digital fabrication in architecture, Autom. Constr. 128 (2021) 103762, https://
4039–4048, https://doi.org/10.1109/TVCG.2020.2999335.
doi.org/10.1016/j.autcon.2021.103762.
[30] R. Gifford, The consequences of living in high-rise buildings, Archit. Sci. Rev. 50
[5] J. Di Pasquale, F. Innella, Y. Bai, Structural concept and solution for hybrid
(2007) 2–17, https://doi.org/10.3763/asre.2007.5002.
modular buildings with removable modules, J. Archit. Eng. 26 (2020) 04020032,
[31] W.F.E. Preiser, E. White, H. Rabinowitz, Post-Occupancy Evaluation (Routledge
https://doi.org/10.1061/(ASCE)AE.1943-5568.0000423.
Revivals), Routledge, London, 2015, https://doi.org/10.4324/9781315713519.
[6] R. Lopez, P.E.D. Love, Design error costs in construction projects, J. Constr. Eng.
[32] F.R. Ammon, Reversing the tide of suburban families? The design, marketing, and
Manag. 138 (2012) 585–593, https://doi.org/10.1061/(ASCE)CO.1943-
occupancy of urban renewal’s high-rise housing, J. Plan. Hist. 19 (2020) 228–255,
7862.0000454.
https://doi.org/10.1177/1538513219897989.
[7] N. Ham, S. Moon, J.-H. Kim, J.-J. Kim, Economic analysis of design errors in BIM-
[33] R. Tokunaga, M. Murota, Residential preferences based on life stage groups of
based high-rise construction projects: case study of Haeundae L project, J. Constr.
residents and factors related to types of intentions to continue living in or
Eng. Manag. 144 (2018) 05018006, https://doi.org/10.1061/(ASCE)CO.1943-
relocating from super-high-rise condominiums, J. Asian Architect. Build. Eng. 22
7862.0001498.
(2023) 765–782, https://doi.org/10.1080/13467581.2022.2052299.
[8] M. Almashaqbeh, K. El-Rayes, Multiobjective model for optimizing the planning of
[34] J. Ghanbari, Psychological role of architecture in social integration of immigrants
floor plans, finishing level, and transportation in modular construction, J. Archit.
in multicultural cities: design experience of a residential mixed-function high-rise,
Eng. 28 (2022) 04022029, https://doi.org/10.1061/(ASCE)AE.1943-
J. Int. Migr. Integr. 20 (2019) 577–592, https://doi.org/10.1007/s12134-018-
5568.0000563.
0616-7.
[9] M. Almashaqbeh, K. El-Rayes, Optimizing the modularization of floor plans in
[35] Z.A. Kiliç, A.K. Yener, Determining proper daylighting design solution for visual
modular construction projects, J. Build. Eng. 39 (2021) 102316, https://doi.org/
comfort and lighting energy efficiency: a case study for high-rise residential
10.1016/j.jobe.2021.102316.

20
J. Liu et al. Automation in Construction 159 (2024) 105264

building, J. Phys. Conf. Ser. (2021) 2069–012156, https://doi.org/10.1088/1742- [58] H. Liu, J.C. Cheng, V.J. Gan, S. Zhou, A novel data-driven framework based on BIM
6596/2069/1/012156. and knowledge graph for automatic model auditing and quantity take-off, Adv.
[36] W. Liao, X. Lu, Y. Huang, Z. Zheng, Y. Lin, Automated structural design of shear Eng. Inform. 54 (2022) 101757, https://doi.org/10.1016/j.aei.2022.101757.
wall residential buildings using generative adversarial networks, Autom. Constr. [59] J. Hao, L. Zhao, J. Milisavljevic-Syed, Z. Ming, Integrating and navigating
132 (2021) 103931, https://doi.org/10.1016/j.autcon.2021.103931. engineering design decision-related knowledge using decision knowledge graph,
[37] J. Schmidt, C.T. Griffin, Barriers to the design and use of cross-laminated timber Adv. Eng. Inform. 50 (2021) 101366, https://doi.org/10.1016/j.aei.2021.101366.
structures in high-rise multi-family housing in the United States, in: Structures and [60] P. Liu, H. Qi, J. Liu, L. Feng, D. Li, J. Guo, Automated clash resolution for
Architecture, CRC Press, 2013, https://doi.org/10.1201/b15267-305. reinforcement steel design in precast concrete wall panels via generative
[38] M. Alavi, D.E. Leidner, Review: knowledge management and knowledge adversarial network and reinforcement learning, Adv. Eng. Inform. 58 (2023)
management systems: conceptual foundations and research issues, MIS Q. 25 102131, https://doi.org/10.1016/j.aei.2023.102131.
(2001) 107–136, https://doi.org/10.2307/3250961. [61] R.E. Oxman, Precedents in design: a computational model for the organization of
[39] D. Zhong, J. Fan, G. Yang, B. Tian, Y. Zhang, Knowledge management of product precedent knowledge, Des. Stud. 15 (1994) 141–157, https://doi.org/10.1016/
design: a requirements-oriented knowledge management framework based on 0142-694X(94)90021-3.
Kansei engineering and knowledge map, Adv. Eng. Inform. 52 (2022) 101541, [62] M. Lou Maher, P.-S. Liew, N. Gu, L. Ding, An agent approach to supporting
https://doi.org/10.1016/j.aei.2022.101541. collaborative design in 3D virtual worlds, Autom. Constr. 14 (2005) 189–195,
[40] J. Jia, Y. Zhang, M. Saad, An approach to capturing and reusing tacit design https://doi.org/10.1016/j.autcon.2004.07.008.
knowledge using relational learning for knowledge graphs, Adv. Eng. Inform. 51 [63] W. Yang, C. Fu, X. Yan, Z. Chen, A knowledge-based system for quality analysis in
(2022) 101505, https://doi.org/10.1016/j.aei.2021.101505. model-based design, J. Intell. Manuf. 31 (2020) 1579–1606, https://doi.org/
[41] D. Zhu, Z. Zhang, L. Shi, J. Qian, S. Qimuge, D. Song, A hierarchical assembly 10.1007/s10845-020-01535-8.
knowledge representation framework and microdevice assembly ontology, Adv. [64] R. Wang, J. Milisavljevic-Syed, L. Guo, Y. Huang, G. Wang, Knowledge-based
Eng. Inform. 53 (2022) 101705, https://doi.org/10.1016/j.aei.2022.101705. design guidance system for cloud-based decision support in the design of complex
[42] V.J.L. Gan, BIM-based graph data model for automatic generative design of engineered systems, J. Mech. Des. 143 (2021) 072001, https://doi.org/10.1115/
modular buildings, Autom. Constr. 134 (2022) 104062, https://doi.org/10.1016/j. 1.4050247.
autcon.2021.104062. [65] H.P.N. Nagarajan, H. Mokhtarian, H. Jafarian, S. Dimassi, S. Bakrani-Balani,
[43] Z. Wang, R. Sacks, T. Yeung, Exploring graph neural networks for semantic A. Hamedi, E. Coatanéa, G. Gary Wang, K.R. Haapala, Knowledge-based design of
enrichment: room type classification, Autom. Constr. 134 (2022) 104039, https:// artificial neural network topology for additive manufacturing process modeling: a
doi.org/10.1016/j.autcon.2021.104039. new approach and case study for fused deposition modeling, J. Mech. Des. 141
[44] N. Skandhakumar, F. Salim, J. Reid, R. Drogemuller, E. Dawson, Graph theory (2018) 021705, https://doi.org/10.1115/1.4042084.
based representation of building information models for access control [66] R. Hu, Z. Huang, Y. Tang, O. Van Kaick, H. Zhang, H. Huang, Graph2Plan: learning
applications, Autom. Constr. 68 (2016) 44–51, https://doi.org/10.1016/j. floorplan generation from layout graphs, ACM Trans. Graph. 39 (2020), https://
autcon.2016.04.001. doi.org/10.1145/3386569.3392391, 118–1.
[45] P. Liu, J. Liu, L. Feng, W. Wu, H. Lan, Automated clash free rebar design in precast [67] M. Aalaei, M. Saadi, M. Rahbar, A. Ekhlassi, Architectural layout generation using
concrete exterior wall via generative adversarial network and multi-agent a graph-constrained conditional generative adversarial network (GAN), Autom.
reinforcement learning, in: T. Ahram (Ed.), Advances in Artificial Intelligence, Constr. 155 (2023) 105053, https://doi.org/10.1016/j.autcon.2023.105053.
Software and Systems Engineering, Springer International Publishing, Cham, 2020, [68] Z. Jiang, H. Wen, F. Han, Y. Tang, Y. Xiong, Data-driven generative design for mass
pp. 546–558, https://doi.org/10.1007/978-3-030-20454-9_54. customization: a case study, Adv. Eng. Inform. 54 (2022) 101786, https://doi.org/
[46] N. Nauata, K.-H. Chang, C.-Y. Cheng, G. Mori, Y. Furukawa, House-GAN: Relational 10.1016/j.aei.2022.101786.
Generative Adversarial Networks for Graph-constrained House Layout Generation, [69] S. Chaillou, ArchiGAN: Artificial intelligence x architecture, in: P.F. Yuan, M. Xie,
2020, https://doi.org/10.48550/arXiv.2003.06988. N. Leach, J. Yao, X. Wang (Eds.), Architectural Intelligence: Selected Papers from
[47] N. Nauata, S. Hosseini, K.-H. Chang, H. Chu, C.-Y. Cheng, Y. Furukawa, House-GAN the 1st International Conference on Computational Design and Robotic Fabrication
++: Generative Adversarial Layout Refinement Networks, 2021, https://doi.org/ (CDRF 2019), Springer Nature, Singapore, 2020, pp. 117–127, https://doi.org/
10.48550/arXiv.2103.02574. 10.1007/978-981-15-6568-7_8.
[48] B. Fu, Y. Gao, W. Wang, Dual generative adversarial networks for automated [70] P. Ghannad, Y.-C. Lee, Automated modular housing design using a module
component layout design of steel frame-brace structures, Autom. Constr. 146 configuration algorithm and a coupled generative adversarial network (CoGAN),
(2023) 104661, https://doi.org/10.1016/j.autcon.2022.104661. Autom. Constr. 139 (2022) 104234, https://doi.org/10.1016/j.
[49] F. Jiang, J. Ma, C.J. Webster, X. Li, V.J.L. Gan, Building layout generation using autcon.2022.104234.
site-embedded GAN model, Autom. Constr. 151 (2023) 104888, https://doi.org/ [71] Z. Fan, J. Liu, L. Wang, G. Cheng, M. Liao, P. Liu, Y.F. Chen, Automated layout of
10.1016/j.autcon.2023.104888. modular high-rise residential buildings based on genetic algorithm, Autom. Constr.
[50] H. Zheng, W. Huang, Understanding and visualizing generative adversarial 152 (2023) 104943, https://doi.org/10.1016/j.autcon.2023.104943.
network in architectural drawings, in: W. Huang, M. Williams, D. Luo, Y.-S. Wu, [72] L. Wang, J. Liu, Y. Zeng, G. Cheng, H. Hu, J. Hu, X. Huang, Automated building
Y. Lin (Eds.), Learning, Prototyping and Adapting - Short Papers Proceedings, The layout generation using deep learning and graph algorithms, Autom. Constr. 154
Association for Computer-Aided Architectural Design Research in Asia (CAADRIA), (2023) 105036, https://doi.org/10.1016/j.autcon.2023.105036.
2018, pp. 233–238. https://scholars.cityu.edu.hk/en/publications/publication [73] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, S. Hochreiter, GANs Trained by
(f5e1a8c8-4231-40d2-bc89-d4c802fb3ce6).html. a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium, 2018,
[51] P. Liu, R. Xiong, P. Tang, Mining Observation and Cognitive Behavior Process https://doi.org/10.48550/arXiv.1706.08500.
Patterns of Bridge Inspectors, Computing in Civil Engineering, 2022, pp. 604–612, [74] I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, Courville, Improved training of
https://doi.org/10.1061/9780784483893.075. wasserstein gans, 2017, https://doi.org/10.48550/arXiv.1704.00028.
[52] R. Xiong, Y. Zhu, Y. Wang, P. Liu, P. Tang, Facilitating construction scene [75] Z. Abu-Aisheh, R. Raveaux, J.-Y. Ramel, P. Martineau, An exact graph edit distance
understanding knowledge sharing and reuse via lifelong site object detection, in: algorithm for solving pattern recognition problems, in: 4th International
L. Karlinsky, T. Michaeli, K. Nishino (Eds.), Computer Vision – ECCV 2022 Conference on Pattern Recognition Applications and Methods 2015, 2015, https://
Workshops, Springer Nature, Switzerland, Cham, 2023, pp. 228–243, https://doi. doi.org/10.5220/0005209202710278 pp. hal-01168816.
org/10.1007/978-3-031-25082-8_15. [76] F. Zhang, N. Nauata, Y. Furukawa, Conv-MPN: Convolutional message passing
[53] P. Liu, J. Xing, R. Xiong, P. Tang, Sharing construction safety inspection neural network for structured outdoor architecture reconstruction, in: In: 2020
experiences and site-specific knowledge through XR-augmented visual assistance, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE,
ISARC Proc. (2022) 21–24, https://doi.org/10.22260/ICRA2022/0008. Seattle, WA, USA, 2020, pp. 2795–2804, https://doi.org/10.1109/
[54] X.-Y. Wang, Y. Yang, K. Zhang, Customization and generation of floor plans based CVPR42600.2020.00287.
on graph transformations, Autom. Constr. 94 (2018) 405–416, https://doi.org/ [77] J. Ohnson, A. Gupta, L. Fei-Fei, Image generation from scene graphs, in:
10.1016/j.autcon.2018.07.017. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
[55] J. Zhu, P. Wu, X. Lei, IFC-graph for facilitating building information access and 2023, pp. 1219–1228, https://doi.org/10.1109/CVPR.2018.00133.
query, Autom. Constr. 148 (2023) 104778, https://doi.org/10.1016/j. [78] Z. Fan, J. Liu, L. Wang, G. Cheng, M. Liao, P. Liu, Y.F. Chen, Automated layout of
autcon.2023.104778. modular high-rise residential buildings based on genetic algorithm, Autom. Constr.
[56] J. Liu, P. Liu, L. Feng, W. Wu, D. Li, Y.F. Chen, Automated clash resolution for 152 (2023) 104943, https://doi.org/10.1016/j.autcon.2023.104943.
reinforcement steel design in concrete frames via Q-learning and building [79] Anon, opencv-python: Wrapper package for OpenCV python bindings, 2023.
information modeling, Autom. Constr. 112 (2020) 103062, https://doi.org/ Available online: https://github.com/opencv/opencv-python (Accessed April 10,
10.1016/j.autcon.2019.103062. 2023).
[57] J. Liu, P. Liu, L. Feng, W. Wu, H. Lan, Automated clash resolution of rebar design in
RC joints using multi-agent reinforcement learning and BIM, (Vol. 36,, IAARC
Publications, 2019, pp. 921–928.

21