Automation in Construction 157 (2024) 105187

Contents lists available at ScienceDirect

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

Review

Generative AI design for building structures

Wenjie Liao a, Xinzheng Lu a, *, Yifan Fei b, Yi Gu b, Yuli Huang a
a
Key Laboratory of Civil Engineering Safety and Durability of Ministry of Education, Tsinghua University, Beijing 100084, China
b
Beijing Engineering Research Center of Steel and Concrete Composite Structures, Tsinghua University, Beijing 100084, China

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

Keywords: Designing building structures presents various challenges, including inefficient design processes, limited data
Building structural design reuse, and the underutilization of previous design experience. Generative artificial intelligence (AI) has emerged
Data feature representation as a powerful tool for learning and creatively using existing data to generate new design ideas. Learning from
Generative AI algorithm
past experiences, this technique can analyze complex structural drawings, combine requirement texts, integrate
Design evaluation
mechanical and empirical knowledge, and create fresh designs. In this paper, a comprehensive review of recent
Intelligent optimization
research and applications of generative AI in building structural design is provided. The focus is on how data is
represented, how intelligent generation algorithms are constructed, methods for evaluating designs, and the
integration of generation and optimization. This review reveals the significant progress generative AI has made
in building structural design, while also highlighting the key challenges and prospects. The goal is to provide a
reference that can help guide the transition towards more intelligent design processes.

1. Introduction such as DALL-E [8] and ChatGPT by OpenAI and AlphaFold by Deep­
Mind [41], have demonstrated the versatility of generative AI in various
The design of building structures is a nuanced task that necessitates fields and have become a cutting-edge area of research.
the blending of empirical and mechanical knowledge. Engineers have Building structural designs rely primarily on drawings that translate
been consistently discovering, developing, and implementing sophisti­ into structured or image data. Engineers complete the design process
cated computer-aided design technologies to streamline the process and based on an architectural design with multiple constraints such as
target efficient and reliable structural designs. compliance and economy. The end-to-end design process of generative
In the 1980s, intelligent design methods were proposed based on AI is consistent with that of building structural design by engineers
expert system algorithms [62]. In the following years, a series of intel­ [104], equipped with powerful learning and generating capabilities to
ligent design methods based on biologically inspired algorithms tackle the intricate puzzles of intelligent design. Therefore, it has
emerged together with the concept of generative design [80,97,102]. become a new research topic.
Advancements in computer technology drove the digitization and Several comprehensive reviews have been undertaken to explore the
automation of building structural designs forward at an unprecedented recent advancements in intelligent technologies based on AI in building
pace. However, expert system algorithms and biologically inspired al­ structural design, analysis, construction, and maintenance (Table 1).
gorithms find themselves grappling with issues in data learning, design These reviews contribute significantly to the understanding of various AI
rule encoding, and design efficiency, which hamper their wider technologies, including machine learning, deep learning, Generative
application. Adversarial Networks (GANs), Generative Pre-trained Transformer
In recent years, artificial intelligence (AI) technologies, notably deep (GPT) models, and Graph Neural Networks (GNNs). By delving into
learning, have made remarkable progress in learning from existing data these topics, these reviews shed light on the potential applications and
and generating new designs, which sets them apart from conventional benefits of these AI technologies in building structures.
expert systems and biologically inspired algorithms. Generative AI In the realm of design and analysis, several notable review studies
design, which harnesses machine learning algorithms to learn from data have been conducted. Aldwaik and Adeli [59] conducted a review
and unfolds new content, has been identified as one of the top ten focusing on the optimization of high-rise building structures, utilizing
technology trends for 2023 [12]. Intelligent generative technologies, nature-based optimization approaches like the neural dynamics model

* Corresponding author.
E-mail address: [email protected] (X. Lu).

https://doi.org/10.1016/j.autcon.2023.105187
Received 11 June 2023; Received in revised form 15 October 2023; Accepted 5 November 2023
Available online 11 November 2023
0926-5805/© 2023 Elsevier B.V. All rights reserved.
W. Liao et al. Automation in Construction 157 (2024) 105187

Table 1 design and visualization within the context of the building industry 4.0.
Scopes of existing review studies related to AI-assistant building structural Wu et al. [11] presented a comprehensive state-of-the-art review on the
design. utilization of GANs to tackle challenging tasks in the built environment.
Dsgn. Anlys. C&O&M DL ML Optim. DRL Saka et al. [9] assessed the potential of GPT models in the construction
(inc. industry, identifying opportunities for their implementation throughout
Gen. the project lifecycle. Jia et al. [110] explored diverse approaches for
AI)
constructing graph data from common construction data types and
Aldwaik & √ √ √ highlighted the significant potential of GNNs for the construction
Adeli [59]
industry.
Chi et al. [37] √ √ √
Amezquita- √ √ √ √ However, there is a noticeable gap in the analysis and discussion
Sanchez surrounding the development and application of generative AI in the
et al. [49] structural design of buildings. This gap results in a lack of systematic and
Li and Adeli √ √ √ targeted research within this field. Upon reviewing existing studies, it
[120]
Afzal et al. √ √ √
becomes evident that the research findings in this area are fragmented
[58] and incomplete. Researchers encounter challenges in addressing specific
Sun et al. [35] √ √ √ √ issues, while the entry barriers for new researchers entering this field are
Pizarro et al. √ √ √ on the rise. These factors collectively impede further progress and
[77]
hinder the advancement of generative AI in building structural design.
Baduge et al. √ √ √ √ √
[95] Therefore, this study aims to provide an overview of existing intelligent
Wu et al. [11] √ √ √ √ structural design methods and focuses on the achievements and appli­
Zakian & √ √ cations of generative AI in building structural design, which has un­
Kaveh [69] dergone rapid development over the past three years. The analysis of
Yüksel et al.
existing studies reveals that the primary research areas in generative AI-
√ √ √ √
[65]
Wang et al. √ √ √ based structural design encompass data representation, generation al­
[19] gorithms, evaluation methods, and the integration of AI generation with
Omrany et al. √ √ √ √ design optimization. By conducting investigations in these pertinent
[34]
domains, the advancement of AI-based building structural design tech­
Saka et al. [9] √ √ √ √
Jia et al. [110] √ √ √ nology can be effectively facilitated. Accordingly, this study will pre­
Topuz & √ √ √ √ √ dominantly center around showcasing the accomplishments of existing
Çakici Alp research within these specific areas.
[14] Section 2 provides a macro-introduction to current AI methods and
Ko et al. [42]
their applications in building structural design. Sections 3–5 review the
√ √ √
Ours √ √ √ √ √ √
data feature representation, generative algorithm construction, and
Dsgn. = Design, Anlys. = Analysis, C&O&M = Construction & Operations & evaluation methods in generative AI. Section 6 introduces methods
Maintenance, DL = Deep Learning, Gen. AI = Generative AI, ML = Machine coupling intelligent generation and optimization, and Section 7 presents
Learning, Optim. = Optimization, and DRL = Deep Reinforcement Learning.
typical engineering application cases of intelligent building structure
design. Section 8 provides relevant conclusions and prospects for the
and genetic algorithms. Chi et al. [37] and Omrany et al. [34] explored future development of this research field.
the integration of building information modeling (BIM) with smart Consequently, this study highlights the potential of generative AI in
technologies in structural design practice, such as structural design, building structural design and the need for further research in this field.
planning, analysis, and optimization. Afzal et al. [58] provided an The development and application of generative AI in structural design
overview of structural components, optimization strategies, and the can significantly enhance the efficiency and accuracy of the design
utilization of various computational tools in RC structural design opti­ process, leading to more sustainable and safer building structures.
mization. Pizarro et al. [77] reviewed rule- and learning-based methods
for intelligent recognition and design in architecture. Sun et al. [35] 2. Introduction to artificial intelligence (AI) methods and
delved into the research on machine learning applications in building applications in building structural design
structure design and performance evaluation. Zakian and Kaveh [69]
presented an overview of seismic design optimization, encompassing 2.1. AI methods
common solution methods, optimization problem types, and optimiza­
tion goals. Yüksel et al. [65] analyzed the research status of genetic al­ Research on AI methods has been conducted for a considerable
gorithms, fuzzy logic, and machine learning methods in engineering period. In this study, the timeline was partitioned using 2012 as a
structure design, covering areas such as design generation, evaluation, reference point, which marks the “deep learning era” owing to the
optimization, decision-making, and modeling. Wang et al. [19] con­ highly effective image classification algorithm AlexNet [4]. According to
ducted a review specifically on the application of AI technology in Yüksel et al. [65], as illustrated in Fig. 1, most AI methods before 2012
material and structural analyses within the field of civil engineering. were categorized as classical AI methods, whereas those after 2012 were
Topuz and Çakici Alp [14] evaluated the current state of machine referred to as modern AI methods. This review focuses primarily on
learning in computer-aided design, engineering, and manufacturing of deep-learning algorithms developed after 2012.
architecture. Lastly, Ko et al. [42] introduced advancements in auto­
mated spatial layout planning combined with AI. (1) Classical AI methods
Several review studies also explored AI applications in the realm of
construction and maintenance. Amezquita-Sanchez et al. [49], and Li Representative methods include knowledge-based systems and bio­
and Adeli [120] provided an insightful review and outlook on the uti­ logically inspired algorithms. Knowledge-based systems include expert
lization of machine learning technologies in structural system identifi­ systems [23], fuzzy logic [48], and generative design grammars [1].
cation, health monitoring, vibration control, design, and optimization. Biologically inspired algorithms include genetic algorithms [31], parti­
Baduge et al. [95] introduced the latest advancements in applying AI cle swarm optimization [84], and cellular automata [25].
technologies, including machine learning and deep learning, in building

2
W. Liao et al. Automation in Construction 157 (2024) 105187

Fig. 1. Artificial intelligence (AI) methods widely adopted in building structural design.

(2) Modern AI methods advancements have provided engineers with enhanced initial design
solutions, elevating both design efficiency and quality. The insights from
Since 2012, the AI field has undergone a comprehensive develop­ these related studies serve as crucial references for the development of
ment driven by deep learning, which can handle big data, extract high- generative AI-based intelligent design in building structures.
dimensional features, and significantly improve learning from data
[111]. Typical deep learning methods include deep convolutional neural 2.2. Application of AI in building structural design
networks (CNNs) [4,111], deep graph neural networks (GNNs) [99], and
deep recurrent neural network/long short-term memory (RNN/LSTM) The building structural design process can be divided into three
[40,90]. In addition, advanced deep learning algorithms have been primary stages: conceptual design (scheme design), detailed design
developed using these methods, such as the variational autoencoder (preliminary or optimization design), and construction drawing design.
(VAE) [24], generative adversarial network (GAN) [38], transformer Among these stages, the conceptual design stage significantly impacts
[10], and diffusion models [28]. In recent years, generative AI the final design outcome and relies heavily on design experience and
[10,28,38,96] has become a representative technological advancement knowledge (as shown in Fig. 2). Therefore, in the conceptual design
in deep learning research. The emergence of generative AI or deep stage, AI plays a significant role in the structural design of buildings.
generative methods has opened new possibilities for intelligent design. The core task of a conceptual building structural design is design
Presently, modern AI methods have found extensive applications in generation (also known as synthesis). According to Maher [62], building
various engineering fields, including architecture, mechanical engi­ structural design synthesis methods mainly refer to the generation of
neering, and aerospace engineering [65]. They have proven to be corresponding designs guided by specific knowledge fragments, such as
effective in generating architectural layouts, rendering architectural heuristic rules and descriptive frameworks. In addition, with the
images [27,39,56,57,63,64,79,89,101,107], generating wheel struc­ development and application of generative AI technologies that can
tures [92], and aerodynamic shapes of aircraft [22]. These learn data features and generate new designs, current structural design

Fig. 2. Influence of the building structural design phase on project cost [15].

3
W. Liao et al. Automation in Construction 157 (2024) 105187

synthesis methods mainly include heuristic search-based, descriptive methods and accounting for 48% of the total research, showing a trend
growth-based, and generative AI learning-based designs (Table 2). of surpassing classical AI methods.
Heuristic search-based designs primarily use biologically inspired In summary, by analyzing the recent trends in intelligent building
computation techniques, such as genetic algorithms, particle swarm structural design, AI-based methods using deep learning have gradually
optimization, and cellular automata [2,5–7,33,43–45,54,61,80, become the primary research focus. These methods are referred to by
88,91,93,112,119]. Descriptive growth-based designs primarily use different names, including intelligent generative design, deep generative
generative design grammar [1,16,26,53,87,108]. These intelligent design, intelligent design, and deep-learning-based design. However,
design methods have been widely applied in multiple areas of building the core technology of these methods primarily involves learning from
structural design, such as multi-scheme search, comparative selection, existing design data, empirical knowledge, and physical principles to
and material optimization, and have promoted the progress of digitali­ master the ability to generate new designs intelligently. As this is
zation and automation in structural design. However, these classical consistent with the essence of generative AI, this study collectively refers
methods face difficulties regarding data learning and design efficiency. to these methods as generative AI-based intelligent designs for building
In contrast, generative learning is an intelligent design method domi­ structures.
nated by generative AI, which has powerful design data learning and
efficient new design generation capabilities and has been continuously 3. Data feature representation and dataset construction
developing and advancing in recent years.
The main architecture of the generative AI algorithms is shown in The representation of the data features is crucial in building struc­
Fig. 3, which includes a dataset, a neural network model, and a loss tural intelligent design. This representation is related to the geometric
function module. The development and application of generative AI in and mechanical characteristics of the structural design and is subject to
building structural design will mainly focus on these three parts, and a the limitations of deep-learning algorithms. In building structural
detailed analysis and summary are presented in Sections 3–5. design, data features mainly include topological, pattern, and size fea­
To further understand the current state of research on intelligent tures. Topological features refer to the layout of structural components
structural design, we conducted a search of the Web of Science Core in space and the connection relationship between components to
Collection for relevant papers. First, a search was conducted on AI-based effectively resist overall lateral and vertical loads. Pattern features refer
structural design methods, yielding 188 search results. Subsequently, to the local geometric configuration of components that must be in
searches were conducted on classical AI-based and modern AI-based special patterns (such as L-shaped, T-shaped, and C-shaped) to resist the
structural design methods, resulting in 130 and 74 search results, load of the local component. Size features refer to the cross-sectional
respectively. The search formulae are listed in Table 3. sizes of structural components.
Fig. 4(a) shows the change in the number of research papers on AI- Generative AI algorithms utilize convolutional neural networks
based building structural designs, indicating a continuous increase in (CNNs) to process data as tensors (including 1-dimensional vectors, 3-
intelligent design research over the years, with a significant increase dimensional images, and n-dimensional matrices). In contrast, graph
since 2020. By comparing the classical and modern AI methods, as neural networks (GNNs) are suitable for processing data in graphs (node
shown in Fig. 4(b), it can be observed that classical AI was the main­ and edge representations, with 1-dimensional vector features embedded
stream research method for intelligent design before 2012. Since 2012, in nodes and edges). Tensor data can better express patterns and size
modern AI has rapidly developed, and the related research papers have features in the design, whereas graphs are more conducive to expressing
increased threefold; however, some differences still exist compared to topological features.
classical AI methods. Since 2020, generative AI technology has signifi­ Therefore, generative AI algorithms can effectively extract features
cantly improved its capabilities and has been developing in-depth in the and learn by appropriately representing the data features in building
field of structural design, with research papers on par with classical AI structural design. This section reviews the data feature representation

Table 2
Comparison of three mainstream intelligent design modes.
Heuristic search-based Descriptive growth-based Generative AI learning-based

By leveraging existing data, AI comprehends the

Utilizing the provided initial solution within The process involves establishing generation rules
mapping patterns between building and structure.
Concept predetermined boundary constraints, the and constraints, commencing from the initial
Trained AI generates a corresponding structural design
description iterative process aims to systematically design, and subsequently iteratively optimizing it
for a new building design in a single comprehensive
search for the optimal result. based on the descriptive generation rules.
step.
pg(x)
pg(x) pdata
pdata(x)(x)

Concept
illustration
Generated
Generated data
data Sample
Sample data
data

Representative
Genetic algorithms Generative design grammar GAN
methods
Initial structural
Required Partial required Not required
design
Manually
Required Required Partial required
defined rules
Iteration design Required Required Not required
Learning Not required Not required Required
Performance High Relatively high Medium
Efficiency Relatively low Medium High

4
W. Liao et al. Automation in Construction 157 (2024) 105187

Fig. 3. Algorithm framework of generative AI.

(1) Tensor-based representation

Table 3
Search objects and formulas used in the Web of Science core collection and the
(2) Graph-based representation
resulting number (accessed on April 20, 2023).
Search objects Search formulas Number of
studies 3.2. Dataset construction

(TS = (“building” OR “architecture*”)

AND TS = (“structur* design”) AND TS =
After determining the data feature representation method, the
(“intelligen*” OR “automat*” OR collected data were processed to construct the dataset. Table 5 provides
“artificial intelligence” OR “design detailed information on the datasets described in the current literature,
Classical AI-based intelligence” OR “generative” OR including the design object, data content, data quantity, and whether
building structural “optimiz*” OR “explorat*”) AND TS = 130
they are open source. Currently, the construction of datasets in building
design (“expert system*” OR “fuzzy logic” OR
“genetic algorithm” OR “generative structural intelligent design is relatively scarce, and the introduction of
grammar” OR “evolution” OR “particle datasets is not sufficiently comprehensive. One significant reason is the
swarm optimization” OR “cellular limited availability of publicly accessible field data.
automata”)) Data augmentation methods are widely used to address the issue of
(TS = (“building” OR “architecture*”)
AND TS = (“structur* design”) AND TS =
limited data. (1) Image data augmentation methods, such as flipping,
(“intelligen*” OR “automat*” OR symmetry, and translation of images, can effectively increase the data up
“artificial intelligence” OR “design to four times the original data without changing the attributes of the
Modern AI-based intelligence” OR “generative” OR structural design [13,104]. Augmenting data through overall image
building structural “optimiz*” OR “explorat*”) AND TS = 74
segmentation can increase the data by hundreds of times by partitioning
design (“machine learning” OR “deep learning”
OR “neural network” OR “generative the completed design image into several local design images without
adversarial network” OR “variational changing the building and structural size [71,76]. (2) Tensor data
autoencoder” OR “transformer” OR augmentation can generate hundreds or thousands of data by parametric
“diffusion model”)) design for structural topology and component size [30,46,72]. (3) Graph
AI-based building
The sum of these two formulas 188 data augmentation can also be performed based on the spatial relative
structural design
coordinate position in each node of the graph, where data can be
augmented >3000 times by overall translation, flipping, and rotating
coordinates [70,74].
methods and dataset construction used in existing research.

3.1. Data feature representation 3.3. Summary of data feature representation

Table 4-1 and Table 4-2 summarize the tensor-based and graph- Currently, there exist significant variations in the extraction and
based data representation methods used in related research, and Fig. 5 representation of data features for different building structure forms.
illustrates the typical representation methods. Tensor-based data rep­ Whether it involves pixel image representation, graph representation, or
resentation is more commonly used in current research, whereas graph- vector representation, it is crucial to employ suitable encoding methods
based representations are less common. Tensor representation is more that effectively capture the topological, geometric, spatial, and other
intuitive and straightforward, and thus it is more widely used in current pertinent information of building structure design. When coupled with
generative AI-based intelligent design methods. Although graphs are corresponding generative AI algorithms, these encoding methods can
better suited for expressing topological features, they require complex facilitate the generation of structural designs. However, it is important
and sophisticated designs to represent pattern and size features. to acknowledge that each representation method possesses its own
limitations and cannot comprehensively capture all design features.
Therefore, it becomes imperative to address the following aspects:

5
W. Liao et al. Automation in Construction 157 (2024) 105187

Fig. 4. Development history of intelligent design of building structures.

Fig. 5. Tensor-based and graph-based representation of data features. (a) Image-based representation [104]. (b) Vector-based representation [78]. (c) Graph-based
representation [55,72].

(1) Achieve a fusion representation of multi-modal data features, (3) Lastly, the rapid advancement of intelligent design in this field
enabling the simultaneous learning of information related to to­ heavily relies on the accumulation and support of open-source
pology, geometry, space, and design conditions. This approach data. Therefore, there is a need for increased sharing of open-
will enhance the comprehensive understanding and representa­ source data within the architectural community to facilitate
tion of building structure designs. further progress and innovation.
(2) Propose a general representation technique that can effectively
densify different types of building structure design features, 4. Generative AI algorithms for structural design
thereby addressing the issue of sparsity associated with archi­
tectural design information. This technique will allow for a more The design methods and rules for the different types of building
complete representation of diverse design components. structures vary significantly. Therefore, the corresponding generative AI

6
W. Liao et al. Automation in Construction 157 (2024) 105187

Fig. 6. Typical generative AI methods for building structural design. (a) GAN [104], (b) GNN [55], (c) VAE [78], (d) ANN [60].

algorithms shall be developed to address the design of structures with topology of beam components, with algorithms primarily based on ANN
different features. Generative AI effectively masters the rules for the and CNN methods. The structural topology is primarily generated by
intelligent generation of building structures by learning from existing inputting the geometric configurations.
design data, mechanical mechanisms, and empirical rules. This section
focuses on introducing generative AI algorithms, specifically for con­
ventional residential building structures, complex spatial building 4.4. Summary of generative AI algorithm
structures, and continuous topology structure designs.
A diverse range of generative AI algorithms exists, with GANs, GNNs,
and VAEs being the most commonly utilized techniques. These algo­
4.1. Residential building structures
rithms have demonstrated their efficacy in designing various types of
structures, including shear walls, frames, and frame-brace structures.
Generative AI is primarily used for designing residential building
However, despite their widespread application, there is still consider­
structures such as shear walls, frames, and frame-brace structures. CNNs
able progress to be made before these technologies reach a level of
and GNNs are the primary generative AI technologies used. Researchers
maturity and are extensively embraced in engineering practice. Moving
have further improved the network architecture by incorporating GANs
forward, numerous challenges need to be addressed in the realm of
and GNNs to learn mechanical mechanisms and empirical rules. Atten­
intelligent generation:
tion and LSTM technologies have also been used in the design-
generation process. A detailed analysis of these techniques is pre­
(1) Presently, the generated structural designs by AI tend to be
sented in Table 6 (See Table 6-1, 6-2 and 6-3).
rough, making it difficult to ensure the rationality of local design
details. For instance, components like columns and shear wall
4.2. Spatial structures elements generated by GANs may exhibit omissions in certain
spatial areas, seriously degrading the mechanical performance of
Spatial structures often have a high degree of freedom and building structures. The primary reason behind this issue lies in
complexity in their shape; therefore, generative AI mainly uses VAE GANs relying on learned high-dimensional probability distribu­
algorithms for intelligent design. The design-generation results can be tions instead of rule-based deterministic design. Consequently,
effectively changed by operating the latent space in the VAE network, important pixel details can be missed during the generation
thereby providing a reference for exploring more structural design process.
schemes. (2) Generative AI has limited learning capabilities when it comes to
various constraints in building structural design rules and me­
4.3. Continuous structures chanical requirements. Although the current algorithms learn
constraints such as symmetry and inter-story drift drifts, they
Generative AI is less commonly used in the topological design of struggle to meet more complex rule requirements and mechanical
continuous building structures. Currently, it is mainly used to design the performance constraints [121].

7
W. Liao et al. Automation in Construction 157 (2024) 105187

Table 4-1 Table 4-2

Tensor-based representation for data features. Graph-based representation of data features.
Research Input features Output features Research Input features Output features

Liao et al. Nodes: frame columns (featured

Nodes: frame columns
[104,105]; Lu Image of architectural Zhao et al. as coordinates)
Image of shear wall Edges: beams (features as 1)
et al. [109]; Zhao components layout, [72] Edges: potential beams (features
components layout and non-beams (features as 0)
et al. [73]; Fei Matrix of design conditions coded for non-beams)
et al. [113] Representation 1: bidirectional
Image of shear wall components graph
Nodes: component endpoints
layout, Matrix of isolator Nodes: architectural component
Liao et al. [103] (featured as coordinates)
Matrix of isolation bearing layout parameters endpoints (features as
Edge: shear wall component
locations coordinates)
(features as component
Image of shear wall components Edge: architectural component
Image of beam length)
Zhao et al. [71] layout, (features as one-hot component
components layout Zhao et al.
Image of architectural space type and endpoint coordinates)
[70,74]
Image of frame-core tube Matrix of cross-sectional Representation 2: one-way graph
Fei et al. [116] components layout, size of frame-core tube Nodes: architectural component Nodes: shear wall component
Matrix of design conditions components (features as one-hot component (features as coordinates)
Image of architectural Two-element vector, type and node coordinates) Edge: node connection
Pizarro et al. components layout, representing thickness Edge: architectural component relationship (features as
[75,76] Vector data of design conditions and length of shear wall (features as one-hot component endpoint coordinates)
and structure-related parameters component type and endpoint coordinates)
Stage 1: Image of Node: structural component
Stage 1: Image of architectural
column components (features as [p1; p2; B; T; L], p1
components layout,
Fu et al. [13] layout, and p2 are member endpoints, B is
Stage 2: Image of architectural Chang & Node: structural component
Stage 2: Image of brace the component type, T is the one-
and column components layout Cheng (features as the one-hot code
components layout hot code of the section size, and L
[55] of the section size)
Output 1: Predicted is the floor number)
coordinates of the four Edge: the connection relationship
128 × 128 × 4 tensor of columns of components
architectural layout, previously Output 2: One-hot Nodes: truss endpoints (features
Ampanavos et al. Edge: connectivity of the truss
arranged columns, pixel encoding of the Hayashi and represent boundary conditions)
[86] endpoints, that is, the
coordinate x value, and pixel predicted three column Ohsaki Edges: potential truss components
existence or absence of truss
coordinates y values types (free-standing, [51] (features include existence,
components
column on corner, or geometric properties, stress state)
column on wall) Stage 1: endpoints of the
Stage 1: architectural image
Reconstructed design Zhang et al. column and shear wall
Stage 2: nodes represent the
Design vector (the initial vector (the designed [20] Stage 2: edges represent the
Danhaive and column and shear wall endpoints
coordinates of nodes controlling coordinates of nodes beam and shear wall
Mueller [78]
the spatial structure shape) controlling the spatial
structure shape)
3D design connectivity in building facade design using diffusion models [39]. On the
3D design connectivity map (The map (the output vector other hand, LLMs find extensive application in text sequence
3D space is divided into grids, dim is 21 × 21 × 21 ×
Rodriguez et al. and each node has 13 potential 13, different values
generation [9,106]. However, a key challenge with LLMs is rep­
[46] connections with surrounding indicate different node resenting building structural design features as the 1-dimensional
nodes; the input vector dim is 21 connections scenario sequence data widely used by LLMs. Once the challenge of data
× 21 × 21 × 13) and the corresponding representation is overcome, it is expected that more research will
structural design)
be conducted in this area.
Five-element input vector, Five-element output
n: number of openings; α: degrees vector, different (4) Currently, there is a scarcity of open-source algorithms for AI-
Mirra & Pugnale
opening rotation; d: opening parameters correspond based building structural design, highlighting the need for
[30]
position; w: opening width; c: to different structural further open communication in the field. Some widely used open-
curvature of the support edges design source code packages in AI design for building structures include
⎡ ⎤ Final density value for U-net [67], pix2pix [68], pix2pixHD [98], and graph-sage [100].
Kallioras & Lagaros T1,1
(d)
⋯ T
(d)
⎢ 1,nex
(f)
⎥ each finite element of
[66] ⎢
Tzk = ⎢
⎥
⎣
⋮ ⋱ ⋮ ⎥
⎦ the initial domain In the future, generative AI technologies in the field of building
T T
(d) (d)
structural design need to become more accurate and adhere to empirical
⋯
ney ,1 ney ,nex
(f) (f) (f)
Nodal values of
independent extended rules to ensure their effectiveness and practicality.
Density values of m fine-solution
multi-scale finite
Huang et al. [60] elements in a coarse-resolution
element method
element 5. Evaluation methods for generative AI design
(EMsFEM) shape
functions
After the intelligent design of building structures is completed, it is
essential to assess the quality of the design results to guide the optimi­
(3) The application of more powerful generative AI technologies, zation of the generative algorithm more effectively and improve the
such as large language models (LLMs) and diffusion models, is subsequent design quality. Relevant methods include evaluation during
limited in intelligent architectural structural design due to their the generative AI training and testing phases. The training phase eval­
relatively short development time and high hardware re­ uation involves constructing a loss function that must ensure that it is
quirements. The diffusion model is presently widely used in differentiable and can be embedded in the computational graph of the
image synthesis domains [8,18,32,83] and is expected to be neural network to guide the learning and optimization of generative AI
gradually explored in building structural design to enhance the effectively. However, the testing phase evaluation has no such re­
effectiveness of intelligent design. Typically, diffusion models strictions, and the feasibility of the design results is measured during this
offer superior image generation quality compared to techniques phase.
like GANs and VAEs. Some studies have already been conducted

8
W. Liao et al. Automation in Construction 157 (2024) 105187

Table 5 Table 6-1

Dataset construction. Generative AI algorithms for residential building structural design.
Research Design Data content Data quantity Open Research Generative AI algorithm Structural
object access system

Image of The generative adversarial network (GAN)

architectural and Liao et al. consists of a generator and a discriminator. The
Liao et al. Shear wall structural 193 building [104] generator comprises an encoder comprising
Yes
[104] structure components and projects Zhao et al. convolutional networks and a decoder
two design [71] consisting of deconvolutional networks. The Shear wall
conditions Zhao et al. discriminator is a multi-scale discriminator components
Image of [73] constructed using convolutional networks. Beam
Pizarro architectural Fei et al. Zhao et al. [73] added a self-attention module components
Shear wall 165 building
et al. [75] components and 30 No [113] to the generator network.
structure projects
[76] design feature Liao et al. Liao et al. [105] improved the encoder of the
vectors [105] generator by using a fusion network that
Image of combines text features and image features.
Steel
architectural GAN with mechanical evaluator enhancement.
Fu et al. frame- 110 building
components, frame No Lu et al. The mechanical evaluator is a ResNet-based Shear wall
[13] brace projects
columns, brace [109] performance evaluation surrogate model for components
structure
layout mechanical properties.
4 fundamental GAN with knowledge evaluator enhancement.
Graph of layout shapes, and 20,000 Fei et al. The knowledge evaluator is a tensor operator Frame-core tube
Zhao et al. Frame
parameters of the layout designs No [116] encoding empirical rules into a differentiable structure
[72] structure
frame beam generated by loss function.
parametric design GAN with the mechanical evaluator and rule
Graph of evaluator enhancement. Seismic isolator
Liao et al.
Zhao et al. Shear wall architectural and 324 building The mechanical evaluator and rule evaluator of shear wall
No [103]
[70,74] structure structural projects are a ResNet-based surrogate model and structure
components layout differentiable tensor operator.
1 fundamental Generator and discriminator.
Structured data: Pizarro et al. Shear wall
Mirra & shape, and 800 Pizarro et al. [75] only adopt the generator,
Shell parameters of [75,76] components
Pugnale designs generated No Pizarro et al. [76] use a GAN.
structure structural
[30] by parametric Two GANs are used for a two-stage design
components
design process. The first stage involved column layout Steel frame-
Fu et al. [13]
2 fundamental design, while the second involved brace layout brace structure
Structured data:
shapes, and 6000 design.
Rodriguez 3D parameters of
and 9000 designs No Generative network: A CNN and LSTM-based
et al. [46] structure structural
generated by generative network.
components
parametric design The design process does not generate the entire
design end-to-end in one go. Instead, it Column
Ampanavos
generates only four columns simultaneously components of
et al. [86]
based on the architectural plan and the frame structure
previous column layout design. The process is
5.1. Construction of loss function repeated iteratively to complete the structural
design.
Graph generative network: a GNN that embeds
Generative AI losses mainly include data, mechanical, and empirical edge features is used, which can simultaneously
rule losses. Data loss mainly refers to the difference between the learn node and edge features.
generated and target designs, and common data difference measurement Input: Graph G = {V , E }, the feature of each
methods such as cross-entropy or mean square error (MSE) can be used node v is xv , the depth of the GNN model is K,
the weight matrix of each layer is
to calculate it. Mechanical and empirical rule losses must be constructed
W(k) (k ∈ [1, K] ), the aggregation function is
based on specific features, and the challenge is to ensure that the error of AGGREGATEk (k ∈ [1, K] ).
the loss function can be backpropagated in a neural network computa­ Output: The embedding vector for each node v: Beam
tion. The mechanical loss is typically evaluated using a neural network Zhao et al.
zv . components of
as a surrogate model for the mechanical performance of the design h0v ←xv , ∀v ∈ V frame structure
[72,74]
for k = 1…K do and shear wall
structure, and this surrogate model must be pre-trained. Constructing an components
for v ∈ V do
empirical rule loss function is challenging, and its core is the construc­ ({ })
hkN (v) ←AGGREGATEk hk− 1
tion of differentiable functions based on specific empirical rules. Owing ( (
u , ∀u ∈ N (v)
))
to the significant differences between various empirical rules, it is hkv ←σ Wk ⋅CONCAT hk− 1 k
v , hN (v)

difficult to form a unified construction paradigm. The empirical rule loss end for
⃦ ⃦
function primarily considers symmetry, the uniformity of cross-sectional ⃦ ⃦
hkv ←hkv /⃦hkv ⃦ , ∀v ∈ V
2
dimensions, and other rules. The details of the loss function construction end for
methods used in relevant research can be found in Table 7-1. zv ←hKv , ∀v ∈ V
Graph generative network: The network
consists of an encoder, feed-forward
5.2. Evaluation methods for structural design results Chang & propagation, and a decoder. The encoder maps The cross-
Cheng the input node features to the embedding space. sectional size of
[55] Feed-forward propagation iteratively updates frame structure
The evaluation of the design results mainly focuses on design simi­ each node’s embedded features. The decoder
larity, mechanical compliance, material consumption, and engineer consists of a multi-layer perceptron (MLP).
experience-based judgment to achieve a more reasonable and accurate Note: The GAN is shown in Fig. 6(a), and the physics-enhanced GNN is shown in
design result evaluation. Evaluating design similarity involves Fig. 6(b).
comparing the consistency between the AI and engineer designs without
involving mechanical performance calculations. Therefore, its evalua­
tion efficiency is high, and it is suitable for evaluating a large number of

9
W. Liao et al. Automation in Construction 157 (2024) 105187

Table 6-2 methods and indicators, as well as the development of more

Generative AI algorithms for spatial structure design. generalizable techniques for constructing differentiable loss
Research Generative AI algorithm Structural functions in the field.
system

Variational autoencoder (VAE) generator Addressing these challenges will pave the way for the establishment
network consisting of an encoder, latent of consistent, comprehensive, and objective evaluation methods and
space, and decoder. indicators in intelligent architectural structural design. This, in turn, will
The encoder and decoder are both artificial enable more effective assessment of AI-generated designs and facilitate
Rodriguez et al. 3D spatial
neural network (ANN) models, and the
[46]
latent space is a feature reduction space
structures, their seamless integration into practical engineering applications.
Mirra & Spatial shell
obtained by extracting the input features
Pugnale [30] structures
through the encoder. Multiple structural 6. Integration of intelligent generation and optimization
design schemes can be generated by
randomly sampling the dimension-reduced
Generative AI has significant advantages in efficiently learning from
features in the latent space.
Based on the conventional VAE, adjustments design data and generating new designs. However, it has limitations in
Danhaive and to the decoder-generated design can be Spatial shell terms of design accuracy owing to its probabilistic nature based on high-
Mueller [78] made by controlling the latent space structures dimensional feature mapping, which is not a deterministic design
features through performance conditions. method. As a result, it can be challenging to accurately satisfy the re­
Note: The VAE is shown in Fig. 6(c). quirements of relevant design specifications in local details. To address
this limitation, an effective method is to combine generative AI with
optimization methods to efficiently converge the design to meet speci­
Table 6-3 fication requirements while satisfying the relevant design experience.
Generative AI algorithms for continuous structures design. Specifically, generative AI is used to generate designs, and optimization
Research Generative AI algorithm Structural algorithms are applied to adjust the generated designs based on the
system specifications. Another approach is using deep reinforcement learning
Kallioras & Deep belief networks to predict the (DRL) to combine generative evaluation, optimization, and learning into
Beam topology
Lagaros [66] parameters of structures an algorithm to achieve better design performance and accuracy.
Structural However, the design efficiency of the approach requires further
Huang et al. [60] ANN
topology
improvement.
Note: The ANN is shown in Fig. 6(d).
6.1. Two-stage design with generative AI and optimization
design test results, primarily for comprehensively evaluating AI capa­
bilities. Mechanical compliance and material consumption analyses rely Chang and Cheng [55] proposed a design method that combined
on mechanical analysis, which consumes more time but yields accurate generative AI with genetic algorithms, where the design results of
results. This method is suitable for evaluating only a small number of generative AI were used as seed solutions. The advantages of this
design results. Engineer experience evaluation takes the longest time, method are illustrated in Fig. 7. The seed solutions generated by the AI
requires reasonable design and evaluation survey forms, invites multiple were closer to the optimal value than those generated by random sam­
engineers for analysis, and has a higher requirement for engineering pling (Advantage 1). Because AI can generate multiple seed solutions
quality. Therefore, the proposed method is suitable for engineering ap­ close to the optimal solution, the quality of optimization by genetic al­
plications. Details of the relevant evaluation methods are listed in gorithms can be further improved, and the convergence of the solution is
Table 7-2. better (Advantage 2). Because AI generates excellent seed solutions, the
number of iterations required to reach the optimal solution level for
5.3. Summary of evaluation method random seed optimization is significantly reduced (Advantage 3).
Fei et al. [115] proposed a self-learning method for the intelligent
Objective, reasonable, and well-generalized evaluation methods are design of building structures involving AI generation, genetic algorithm
crucial for measuring technological advancements in the field of intel­ optimization, and AI relearning. This method addresses the challenges of
ligent building structural design. However, certain challenges need to be small data sizes and significant long-tail distributions in the intelligent
addressed: building structure design. Specifically, the method first learns the design
rules based on existing small data samples using generative AI. After
(1) In recent years, evaluation methods and indicators in this domain training, generative AI can generate multiple design results. Subse­
have exhibited inconsistency and incompleteness. The existing quently, the genetic algorithm optimizes the newly generated designs to
evaluation criteria often struggle to comprehensively and objec­ obtain additional high-quality learning samples, enlarging the original
tively reflect the mechanical performance, adherence to empir­ dataset. Finally, the updated dataset was used for AI learning to improve
ical rules, and economic efficiency of structural design. The design quality. It is worth noting that, during the optimization process,
rationality and rigor of evaluation indicators play a crucial role in the design of the shear wall component cross-sectional size can be
determining the practical applicability of AI-based intelligent facilitated by the work of Feng et al. [118]. Furthermore, the quantifi­
design outcomes in engineering practice. Moreover, encoding cation of construction material can be achieved through the imple­
these complex evaluation criteria into executable computer code mentation of an AI-predicting method proposed by Fei et al. [114].
poses difficulties, hampering automated evaluation processes.
(2) AI model training requires the use of differentiable loss functions 6.2. Design generation and optimization based on deep reinforcement
within the framework of deep learning. However, the calculation learning (DRL)
processes for indicators like mechanical performance, adherence
to empirical rules, and economic efficiency in building structural DRL is a self-learning technology that consists of an agent and an
design are generally non-differentiable. This poses a challenge in environment. It is characterized by the interaction between the agent
effectively incorporating domain knowledge into loss functions to and the environment, with the agent learning from the feedback pro­
guide generative AI learning. Hence, there is an urgent need for vided by the environment [85]. As shown in Fig. 8, the DRL process can
research focusing on universal and objective AI design evaluation be described as follows: (1) at each time step, the agent interacts with

10
W. Liao et al. Automation in Construction 157 (2024) 105187

Table 7-1
Loss function (data loss, mechanical loss, empirical loss).
Research Loss function Structural system

Image loss of discriminator:

L D = − Ex∼pdata (x) [log(D(s, x) ) ] − Ez∼pz (z) [log(1 − D(G(s, z) ) ) ]
Liao et al. [104,105]
Image loss of generator:
Lu et al. [109] Shear wall structure;
L GAN = Ez∼pz (z) [log(1 − D(G(s, z) ) ) ]
Fei et al. [113] ∑ [⃦ ⃦ ] Beam structural
L FM = Ex∼pdata (x) Ti=1 ⃦Dk (s, x) − Dk (s, G(s) ) ⃦
⃦ (i) ⃦
Zhao et al. [71,73] components
(i)
1
Fu et al. [13] ∑LenVGG [⃦ ⃦ ]
L VGG = Ex∼pdata (x) j=1 ωj ⃦VGGj (x) − VGGj (G(s) ) ⃦1
x is label, s is input, z is generated data, D() is the discriminative network, G() is the generative network, VGG() is the VGG19 network
Text loss of generator:
( ) vtxt ⋅vtxt
Liao et al. [105] L word = cossim vtxt txt
real , vfake =
(⃦ real⃦ fake
⃦ ⃦ ) Shear wall structure
⃦ ⃦
max ⃦vtxt ⃦ ⋅⃦vtxt ⃦
real 2 fake
2
vtxtis the real text data feature,
real is the fake text data feature, cossim() is cosine similarity
vtxt
fake
Mechanical loss of generator:
⎧
⎨ 1 − ddrift
⎪ drift drift drift
max /dlimit , dmax ≤ dlimit
pPHY = pdrift = ( drift drift )0.5 ,
⎩ dmax /dlimit − 1
⎪ , ddrift drift
max > dlimit

Lu et al. [109] L PHY = PHYE(z) Shear wall structure

ddrift drift
max is the maximum inter-story drift ratio, dlimit is the code limitation of inter-story drift ratio
pPHY is used to evaluate the structural mechanical performance, serving as the label value for training the mechanical performance
surrogate model (PHYE) in the field of intelligent design for building structures. The trained PHYE model outputs the value of
mechanical loss L PHY .
∑⃒ ⃒
L knwl,1 = ⃒Colsym − Colsym ⃒,
∑⃒⃒ ⃒
⃒
L knwl,2 = ⃒Wal − WalDirect,mean ⃒
Ext Ext
Frame-core tube
Fei et al. [116] ∑⃒⃒ ⃒
L knwl,3 = ⃒WalInt − WalInt
⃒ structure
Line,mean ⃒
∑( High )
L knwl,4 = Ele − Ele Low
, if EleHigh > EleLow

Pizarro et al. [75,76] 1 ∑ N ( ) 2 Shear wall structure

L MSE = yi − ̂
yi
N i=1
L = ωcoords L MAE + ωclass L CE
1 ∑N ⃒⃒ ⃒
L MAE = yi − ̂
yi ⃒
N i=1 ( )
⎛ ⎞
Ampanavos et al. exp yn,c
1 ∑N 1 ∑C ⎝ Frame structure
[86] L CE = − log ( y
)̂ n,c
⎠
N n=1 C c=1 ∑C
i=1 exp yn,i

The loss function was defined as the weighted sum of the mean absolute error of the coordinates output layer and the categorical
cross-entropy of the column type output layer, with weights 1.0 and 0.2.
Zhao et al. [72] 1 ∑N ( ( ) )
Frame structure
L = − yi logp(xi ) + 1 − yi log(1 − p(xi ) )
N i=1
1 ∑ N ⃒ ⃒
⃒yi − pi ⃒
L Edge =
N ( i )/
∑N ∑M
Zhao et al. [74] L Node = − i c=1 yic log(pic ) N Shear wall structure
where N is the edge (node) sample number; yi is the target value of the i-th
sample; pi is the predicted value of the i-th sample.
Mechanical loss:
L = ω0 lobj + ω1 ldr + ω2 lvar + ω3 lH
lobj is mass objective;
Chang & Cheng [55] Frame structure
ldr is drift ratio constraint, ldr = Mean{LeakyReLU(|dri − lim|) } ≤ 0;
lvar is variety constraint, lvar = 1 − SumTop6(p) = 0;
lH is entropy constraint, lH = Mean{Hi }/Hmax − α = 0, α = 0.6
z ∼ Enc(x) = qϕ (z|x), ̃ x ∼ Dec(z) = pθ (x|z)
(( )2 )
1 ∑N ( )
L VAE (z; θ, ϕ) = xi − ̃
xi + DKL qϕ (z|x) ‖ pθ (z)
N i=1
Danhaive and Spatial shell
Mueller [78] qϕ projects a high-dimensional input x to a compressed representation z, which the decoder pθ projects back to its original structures
representation as best as possible. θ and ϕ denote their respective network’s weights which are optimized during training to
minimize a composite loss function combining the mean square error (MSE) (reconstruction term) and Kullback–Leibler divergence
(which can be seen as a regularization term)
( ( ) )
Rodriguez et al. [46] 1 ∑N ( )
3D spatial structure
L VAE (z; θ, ϕ) = − ̃ xi log(1 − p(xi ) ) + DKL qϕ (z|x) ‖ pθ (z)
xi logp(xi ) + 1 − ̃
N i=1

L VAE (z; θ, ϕ) = L pixel

like + L prior
Mirra & Pugnale pixel
L = − Eq(z|x) [logp(z|x) ] 3D spatial structure
[30] like
( )
L prior = DKL qϕ (z|x) ‖ pθ (z)
L = ωMSE L MSE + ωPHY L PHY
1 ∑N ( )2
L MSE = yi − ̂yi
N i=1
1 ∑ M
Huang et al. [60] L PHY = (si − ̂s i )2 Structural topology
M i=1
The loss function contains two parts: one is the MSE between the prediction (yi ) and true outputs (̂ y i ); the other one is the mean
square error between the stiffness matrix (si ) calculated by the predicted shape functions and the exact stiffness matrix (̂s i ) obtained
by the EMsFEM of the coarse-resolution elements (the second part is actually a physical constraint).

11
W. Liao et al. Automation in Construction 157 (2024) 105187

Table 7-2
Evaluation methods for design results.
Research Evaluation method Indicator Structural system

SIoU = ηSWratio IoU

⃒ GAN ⃒
⃒R − RENG ⃒
Liao et al. [104,105] ηSWratio = SW ENG SW
Lu et al. [109] RSW
Data similarity analysis, that is, intersection over union (IoU)
Zhao et al. [71,73] Aswall Shear wall structure;
RSW =
Fei et al. [113] Aswall + Ainwall Beam structural
Pizarro et al. Ainter components
IoU =
[75,76] AGAN + AENG − Ainter
Structural natural frequency, inter-story drift ratio, axial
Zhao et al. [74] Mechanical compliance
compression ratio, etc.
Economic feasibility Concrete and steel material consumption
Fei et al. [113] Proportion of cantilevered floor Aminus Shear wall structure
SfloorA = 1 −
Afloor
1 ∑Nex +Nnonex NF
SEP− 1 =
Nex + Nnonex i NF + NT
( )
1 ∑Nex 1 ∑Nimg
Engineer evaluation: SEP− 2 = ηex S j +
Nex i Nimg j
Liao et al. [104] Real or fake judgment, ( ) Shear wall structure
Design rationality 1 ∑Nnonex 1 ∑Nimg
ηnonex Sj
Nnonex i Nimg j

σnonex /μnonex
ηex = ,η = 1 − ηex
σex /μex + σnonex /μnonex nonex
Differences of material consumption Difference between column and brace, respectively
Frame-braced
Fu et al. [13] Engineer grading: Arrangement on key nodes
structure
Rationality of key nodes Reasonability of overall layout
Ampanavos et al.
Evaluation of the accuracy of column layout Accuracy Frame structure
[86]
Zhao et al. [72] Evaluation of the accuracy of beam layout Accuracy, F1 score Frame structure
Accuracy of column cross-sectional size;
Chang & Cheng [55] Accuracy, code limitation of inter-story drift ratio Frame structure
inter-story drift ratio
Danhaive and Spatial shell
Material consumption Structural mass
Mueller [78] structures
(1) Low variance, that is, the presence of several samples that
are similar to each other; and 3D Spatial shell
Mirra & Pugnale [30] No specific indicators, engineers select them
(2) Consistency, that is, the recurrence of some underlying structures
features in all the samples.
(1) Optimized structural compliances
Huang et al. [60] Computational methods for topology optimization Structural topology
(2) Efficiency

design generation and optimization for concrete structures are the to­
pological layout, cross-sectional dimensions of reinforced concrete (RC)
components, and reinforcement design. Jeong and Jo [47] applied DRL
to optimize the reinforcement design of RC frame structures. The agent’s
action network model predicts subsequent design actions based on the
current state of the RC components, and the agent’s neural network
continually learns during the process, ultimately enabling the agent to
complete the RC component reinforcement design quickly. Cheng et al.
[29] performed an optimal design of a shear wall structure layout based
on DRL using a convolutional neural network model to evaluate the
current state of the structure and subsequently execute optimization
actions.
Hayashi and Ohsaki [50,52], Zhu et al. [94], and Kupwiwat et al.
[21] proposed steel truss structure designs based on DRL by combining
GNNs and reinforcement learning. The intelligent design approach
Fig. 7. Advantage of using generative AI-designed results as seed solutions in
characterizes the truss structure as a graph with information such as
the field of intelligent design for building structures [55].
position coordinates and component cross-sectional dimensions repre­
sented as features in the graph. This facilitates the agent’s observation of
the environment and receives a high-dimensional observation, which is
the current state, enabling it to execute actions accordingly. This method
used to obtain a specific state representation using deep learning
has been used to achieve cross-sectional dimension optimization for
methods; (2) based on the expected return, the value function of each
steel frame structures and topological generation and optimization
action is evaluated, and a specific policy is used to map the current state
design for steel truss structures. Fig. 9 shows an example of the structural
to the corresponding action; and (3) the agent takes action, and the
topology design. Similarly, Luo et al. [81,82] represented truss layout
environment responds by providing a reward signal, which is used to
generation and optimization as Markov decision processes using node
update the agent’s policy and value function. The next observation is
positions as optimization variables and a Monte Carlo tree search
made, and the process continues iteratively to optimize the agent’s
(MCTS) to perform truss structure layout design. To address the problem
behavior and achieve a specific goal. DRL has been adopted for the
of exploding action space dimensions, kernel regression was employed
intelligent design of building structures based on the concept of (deep)
to constrain the action space, effectively improving the optimization
reinforcement learning [21,29,36,47,50,52,81,82,94].
efficiency.
In the intelligent design of building structures, the main objects of

12
W. Liao et al. Automation in Construction 157 (2024) 105187

Fig. 8. Concept of deep reinforcement learning (DRL) in intelligent structural design [47].

Fig. 9. Topological generation and optimization of steel truss structures based on DRL [94].

6.3. Summary of integration of generation and optimization 7. Application of generative AI-based design for building
structures
Currently, AI-generated structural designs encounter challenges in
achieving optimal overall design while considering local design aspects. 7.1. Development of intelligent structural design systems
The data used for AI learning is primarily based on engineers’ experience
and may not necessarily represent the ideal design outcomes. Therefore, A cloud design system can be constructed based on generative AI
further algorithmic optimization is required to enhance the quality of AI- design methods for building structures [113]. As illustrated in Fig. 10(a),
generated designs. However, there exists a gap between AI generation the users can create the design project of the shear wall layout, beam
and design optimization: layout, and cross-sectional size of the frame-core tube structural com­
ponents. Subsequently, generative AI can complete the corresponding
(1) The data format generated by AI differs from the mechanical structural scheme design by inputting the architectural scheme and
modeling data format necessary for optimization calculations, design conditions as required (Fig. 10(b)). The structural scheme can
necessitating better coordination in transferring data between then be imported into CAD drawing and structural analysis software
these two stages. (such as Autodesk CAD, PKPM, or YJK) using a parametric modeling
(2) Deep reinforcement learning and conventional optimization al­ system. This allows for the efficient completion of structural scheme
gorithms entail extensive search spaces and action spaces, verification and construction design. Therefore, the design process, from
resulting in lengthy optimization processes and difficulties in the architectural scheme to the preliminary construction drawing, can
effectively meeting the requirements of the design stage. Addi­ be accomplished within 20 min, thus significantly improving the design
tionally, the optimization results may not accurately reflect the efficiency.
design experience of engineers, making it challenging to directly
apply current optimization designs in complex real-world
projects. 7.2. Application and cases
(3) It is essential to establish a generative-optimization-relearning
pathway, where generative AI and optimization algorithms Feng et al. [117] proposed an intelligent design method for shear
collaborate to generate designs, optimize them, and continuously wall structure schemes that met the requirements of project bidding,
learn from the optimized results. This iterative process allows for including shear wall layout design and optimization, component cross-
automatic improvement and refinement. sectional design, and automated construction of structural analysis
models. This study achieved an intelligent and highly automated pro­
Hence, future research needs to address these challenges to strike a cess, from architectural to structural design drawings, structural analysis
balance between design accuracy and efficiency. models, and material consumption statistics (Fig. 11). The method was
applied to a project bidding process, and the results indicated that, for a
single building, the generative AI-based structural design efficiency was
nearly 10–15 times higher than that of manual design. Moreover, the
intelligent design method can simultaneously conduct multiple project

13
W. Liao et al. Automation in Construction 157 (2024) 105187

Fig. 10. Intelligent structural design system. (a) Design platform; (b) typical design outcomes of intelligent design.

Fig. 11. Application of generative AI-based design for building structure [117].

designs, with the structural mechanical performance of the intelligent obtained through long-term optimization.
design meeting code requirements and the material usage almost iden­
tical to that of the engineer-designed structure. 8. Discussion
Additionally, Chang & Cheng [55] conducted a case study focusing
on the rapid design of component cross-sectional dimensions for 8.1. Level of generative AI design for building structure
different steel frame structures. They employed the GNN-based Neu­
ralSim and NeuralSizer modules proposed in their research. Fig. 12(a) Widespread studies and applications of generative AI in the struc­
showcases multiple design options, with a training time of approxi­ tural design of buildings have led to significant progress in this field. To
mately 2.5 h. However, once trained, the NeuralSizer module can better understand the current research status and future development
generate a new design in just 10.07 milliseconds, effectively incorpo­ potential, it is suggested that the degree of generative AI involvement be
rating the constraints of NeuralSim and demonstrating the algorithm’s divided into hierarchical levels from L0 to L5.
generalization capability. Another noteworthy study by Jeong & Jo [47] The L0 level (completely manual design) refers to a structural design
utilized deep reinforcement learning to generate and optimize frame entirely controlled by engineers. The design process depends entirely on
beam design. They applied this method to a two-dimensional, three- human expertise and experience at this level.
story, three-span reinforced concrete frame structure, as depicted in The L1 level (human-led, AI-assisted) refers to the AI providing
Fig. 12(b). In comparison to genetic algorithms (GA) [17] and the BB-BC assistance, such as parameter optimization and rule checking for struc­
algorithm [3], Jeong and Jo [47] proposed method requires minimal tural design. This relatively mature research and application area in­
iteration, effectively learns from existing design data, significantly en­ cludes parametric modeling, intelligent optimization, and automated
hances efficiency, and achieves material usage comparable to results rule checking. The L1 level is the current general level of intelligent

14
W. Liao et al. Automation in Construction 157 (2024) 105187

Fig. 12. Typical generative AI-based design cases. (a) Design case of a steel frame structure [55]. (b) Design case of a concrete frame structure [47].

design, with AI assisting engineers during the design process. 8.2. Challenges
The L2 level (AI-led, human-assisted at certain stages) refers to AI
taking over some of the tasks in certain stages of structural design; A review of existing research on generative AI in building structural
however, it still requires human engineers to check the design results design reveals that although there have been significant advancements,
and complete other design tasks. For example, AI can complete the several critical challenges still need to be addressed. These challenges
structural scheme design, and with the development and application of include but are not limited to
generative AI in structural design, building structural intelligent design
has begun to enter the L2 level. • Limited data availability: Obtaining high-quality drawings for
The L3 level (AI-led, human-assisted for specific projects) refers to training generative AI models is often difficult. This presents a
the ability of AI to complete the entire structural design process of challenge in identifying key features and discovering potential
specific building projects. design rules for small samples.
The L4 level (AI-led, human-assisted for most projects) refers to AI • Sparse feature identification: Only a small proportion of the effective
being able to complete the entire design task for most building struc­ information on the drawing is related to the structural design.
tures. At this level, AI can handle most design processes with minimal Therefore, it is essential to accurately identify, extract, and learn
human intervention, significantly reducing the time and cost required sparse key features to generate high-quality designs.
for design. • Multiple constraints: Building structural design involves various
The L5 level (AI-led for all projects) refers to AI being able to com­ constraints, including spatial layout, mechanical principles, and
plete all structural design tasks for any building project. This is the ul­ empirical rules. It is crucial to construct a neural network model that
timate goal of generative AI-based intelligent design, in which AI can considers these constraints and ensures that the constraint-based loss
handle all design aspects. function is differentiable.
Although progress has been made in the development of levels L1 • Complex representation: Building structural design involves the
and L2, the development of levels L3–L5 remains in its early stages and multimodal data representation of drawings, regulation texts,
requires further breakthroughs in AI. The development of generative AI- graphs, etc. To generate high-quality designs, information must be
based intelligent design has the potential to revolutionize the building extracted from heterogeneous sources to represent multidimensional
structural design process, significantly improving efficiency, accuracy, information.
and cost-effectiveness.

15
W. Liao et al. Automation in Construction 157 (2024) 105187

• Strict indicators: AI-designed structures must satisfy economic in­ evaluation.

dicators; otherwise, they may not be feasible for construction. Furthermore, an effective combination of generative and optimiza­
Therefore, it is essential to integrate generative AI algorithms with tion methods can further integrate their advantages with the potential to
multi-objective and multi-scale optimization algorithms to ensure significantly reduce the optimization search space, improve the opti­
the designs satisfy the required economic indicators. mization efficiency, and enhance the optimization quality. In addition,
• Low fault tolerance: structural design cannot lead to safety errors. combining generative and optimization methods with DRL can learn
Therefore, it is crucial to develop a human-in-loop evaluation strat­ through interaction with the environment, effectively reducing the time
egy that integrates subjective perceptions with objective indicators required for optimization and providing new implementation ap­
to ensure that the generated designs are safe and reliable. proaches for intelligent design.
Based on current generative AI design methods, the research results
8.3. Future outlooks have been applied to create new design paradigms, including intelligent
lightweight structural designs based on cloud platforms, fast material
As intelligent design research in the field of building structures consumption estimation in bidding, and efficient scheme design para­
continues to evolve, it is expected that the corresponding challenges will digms based on the generation–optimization–analysis process.
be addressed. This research outlook can be divided into two categories: In the future, with the development of more powerful intelligent
technical and applicable. algorithms and larger design datasets, machine intelligence will better
understand the theory of building structural design. Better design results
(1) From a technical perspective, the following research re­ and higher levels of automation and intelligence are expected for con­
quirements should be addressed: ventional building structures, whereas AI can improve certain aspects of
the design process for complex building structures. The continued
• Comprehensive representation of multimodal data features and development of generative AI-based intelligent design methods for
feature fusion learning to improve data representation and dataset building structures holds great promise for the construction industry,
construction; the emergence of more open-source data to improve enabling engineers to design safer, more efficient, and more cost-
data availability. effective structures.
• Advancement of generative algorithms such as the diffusion model
and large language model (LLM) (e.g., OpenAI’s ChatGPT and GPT-
4) to gain a deeper understanding of structural design knowledge. Declaration of Competing Interest
• Development of universal and differentiable knowledge loss func­
tions for design evaluation and optimization, improvement of the The authors declare that they have no known competing financial
efficiency of the design optimization process through efficient and interests or personal relationships that could have appeared to influence
precise evaluation methods, and methods to reduce the action space the work reported in this paper.
in DRL.
Data availability
(2) From an application perspective, the outlook for an intelligent
design of building structures involves software, algorithms, and No data was used for the research described in the article.
application foundations. To achieve intelligent design more
effectively in the future, the following steps should be taken: Acknowledgments

• Establish a digital software foundation for the design generation and This work was supported by the Science and Technology Program of
analysis optimization process, using advanced digital technologies the Ministry of Housing and Urban-Rural Development of the People’s
such as building information modeling (BIM). Republic of China (2022-K-073), the China Postdoctoral Science Foun­
• New intelligent algorithms should be actively adopted and combined dation (2022M721879), the Tencent Foundation through the XPLORER
with structural design knowledge to establish an algorithmic foun­ PRIZE, and the Shuimu Tsinghua Scholar Program (2022SM005).
dation with embedded domain knowledge.
• The industry should be encouraged to actively explore the applica­ References
tion of intelligent design and establish an application foundation.
[1] A. Chakrabarti, K. Shea, R. Stone, J. Cagan, M. Campbell, N.V. Hernandez, K.
9. Conclusions L. Wood, Computer-based design synthesis research: an overview, J. Comput. Inf.
Sci. Eng. 11 (2) (2011), 021003, https://doi.org/10.1115/1.3593409.
[2] A. Kanyilmaz, P.R.N. Tichell, D. Loiacono, A genetic algorithm tool for
Generative AI-based intelligent design is an essential approach for conceptual structural design with cost and embodied carbon optimization, Eng.
overcoming the challenges faced by the current structural design in­ Appl. Artif. Intel. 112 (2022), 104711, https://doi.org/10.1016/j.
engappai.2022.104711.
dustry, such as low design efficiency, insufficient data reuse, and diffi­ [3] A. Kaveh, O. Sabzi, Optimal design of reinforced concrete frames using big bang-
culty in the inheriting design experience. This method learns from big crunch algorithm, Int. J. Civil Eng. 10 (3) (2012) 189–200. https://www.
existing design data, understands the rules for generating structural researchgate.net/publication/287868723.
[4] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep
designs based on architectural design, effectively learns the constraints convolutional neural networks, Commun. ACM 60 (6) (2017) 84–90, https://doi.
of drawing and text data, as well as mechanical and empirical knowl­ org/10.1145/3065386.
edge, and efficiently maps to generate designs. Therefore, in recent [5] A. Nimtawat, P. Nanakorn, A genetic algorithm for beam-slab layout design of
rectilinear floors, Eng. Struct. 32 (11) (2010) 3488–3500, https://doi.org/
years, the research focus on intelligent design for building structures has 10.1016/j.engstruct.2010.07.018.
shifted towards generative AI design based on deep learning. [6] A. Nimtawat, P. Nanakorn, Automated layout design of beam-slab floors using a
Research on generative AI design for building structures has mainly genetic algorithm, Comput. Struct. 87 (21− 22) (2009) 1308–1330, https://doi.
org/10.1016/j.compstruc.2009.06.007.
focused on data feature representation and learning, intelligent gener­
[7] A. Nimtawat, P. Nanakorn, Simple particle swarm optimization for solving beam-
ation of structural design, and effective evaluation of design results. slab layout design problems, Procedia Eng. 14 (2011) 1392–1398, https://doi.
Related research has focused on data feature representations of images org/10.1016/j.proeng.2011.07.175.
and graphs. Intelligent generation algorithms have included GAN, GNN, [8] A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, I. Sutskever,
Zero-shot text-to-image generation, in: Proceedings of the 38th International
VAE, and ANN. The design evaluation has included Turing tests, data Conference on Machine Learning, 2021, pp. 8821–8831, https://doi.org/
similarity measurements, mechanical compliance, and economic 10.48550/arXiv.2102.12092.

16
W. Liao et al. Automation in Construction 157 (2024) 105187

[9] A. Saka, R. Taiwo, N. Saka, B. Salami, S. Ajayi, K. Akande, H. Kazemi, GPT models Comput Methods Eng. 22 (1) (2015) 135–151, https://doi.org/10.1007/s11831-
in construction industry: opportunities, limitations, and a use case validation, 014-9127-7.
arXiv (2023) preprint arXiv:2305.18997, https://arxiv.org/abs/2305.18997. [38] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair,
[10] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, Ł. Kaiser, A. Courville, Y. Bengio, Generative adversarial networks, Commun. ACM 63 (11)
I. Polosukhin, Attention is all you need, in: Advances in Neural Information (2020) 139–144, https://doi.org/10.1145/3422622.
Processing Systems 30, 2017, https://doi.org/10.48550/arXiv.1706.03762. [39] J. Chen, Z. Shao, B. Hu, Generating interior design from text: a new diffusion
[11] A.N. Wu, R. Stouffs, F. Biljecki, Generative adversarial networks in the built model-based method for efficient creative design, Buildings 13 (7) (2023) 1861,
environment: a comprehensive review of the application of GANs across data https://doi.org/10.3390/buildings13071861.
types and scales, Build. Environ. (2022), 109477, https://doi.org/10.1016/j. [40] J. Chung, C. Gulcehre, K. Cho, Y. Bengio, Empirical evaluation of gated recurrent
buildenv.2022.109477. neural networks on sequence modeling, arXiv (2014), https://doi.org/10.48550/
[12] Alibaba DAMO Academy, Top ten Technology Trends of DAMO Academy. arXiv.1412.3555 preprint arXiv:1412.3555.
https://damo.alibaba.com/techtrends/2023, 2023 (in Chinese). [41] J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger,
[13] B. Fu, Y. Gao, W. Wang, Dual generative adversarial networks for automated K. Tunyasuvunakool, R. Bates, A. Žídek, A. Potapenko, A. Bridgland, C. Meyer, S.
component layout design of steel frame-brace structures, Autom. Constr. 146 A.A. Kohl, A.J. Ballard, A. Cowie, B. Romera-Paredes, S. Nikolov, R. Jain,
(2023), 104661, https://doi.org/10.1016/j.autcon.2022.104661. J. Adler, T. Back, S. Petersen, D. Reiman, E. Clancy, M. Zielinski, M. Steinegger,
[14] B. Topuz, N. Çakici Alp, Machine learning in architecture, Autom. Constr. 154 M. Pacholska, T. Berghammer, S. Bodenstein, D. Silver, O. Vinyals, A.W. Senior,
(2023) 105012, https://doi.org/10.1016/j.autcon.2023.105012. K. Kavukcuoglu, P. Kohli, D. Hassabis, Highly accurate protein structure
[15] B.C. Paulson, Designing to reduce construction costs, J. Constr. Div. 102 (4) prediction with AlphaFold, Nature 596 (7873) (2021) 583–589, https://doi.org/
(1976) 587–592, https://doi.org/10.1061/JCCEAZ.0000639. 10.1038/s41586-021-03819-2.
[16] C. Königseder, K. Shea, Systematic rule analysis of generative design grammars, [42] J. Ko, B. Ennemoser, W. Yoo, W. Yan, M.J. Clayton, Architectural spatial layout
AI EDAM 28 (3) (2014) 227–238, https://doi.org/10.1017/ planning using artificial intelligence, Autom. Constr. 154 (2023), 105019,
S0890060414000195. https://doi.org/10.1016/j.autcon.2023.105019.
[17] C. Lee, J. Ahn, Flexural design of reinforced concrete frames by genetic algorithm, [43] J. Reisinger, M. Knoll, I. Kovacic, Parametric structural design for automated
J. Struct. Eng. 129 (6) (2003) 762–774, https://doi.org/10.1061/(ASCE)0733- multi-objective optimization of flexible industrial buildings, in: Proceedings of
9445(2003)129:6(762). the International Symposium on Automation and Robotics in Construction Vol.
[18] C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E.L. Denton, K. Ghasemipour, 37, IAARC Publications, 2020, pp. 185–192. https://www.iaarc.org/publicatio
R. Gontijo Lopes, B. Karagol Ayan, T. Salimans, Photorealistic text-to-image ns/fulltext/ISARC_2020_Paper_144.pdf.
diffusion models with deep language understanding, Adv. Neural Inf. Proces. [44] J. Reisinger, M.A. Zahlbruckner, I. Kovacic, P. Kan, X. Wang-Sukalia,
Syst. 35 (2022) 36479–36494, https://doi.org/10.48550/arXiv.2205.11487. H. Kaufmann, Integrated multi-objective evolutionary optimization of production
[19] C. Wang, L.H. Song, Z. Yuan, J.S. Fan, State-of-the-art AI-based computational layout scenarios for parametric structural design of flexible industrial buildings,
analysis in civil engineering, J. Ind. Inf. Integr. 33 (2023), 100470, https://doi. J. Build. Eng. 46 (2022), https://doi.org/10.1016/j.jobe.2021.103766.
org/10.1016/j.jii.2023.100470. [45] J. Reisinger, S. Kugler, I. Kovacic, M. Knoll, Parametric optimization and decision
[20] C. Zhang, M.X. Tao, C. Wang, J.S. Fan, End-to-end generation of structural support model framework for life cycle cost analysis and life cycle assessment of
topology for complex architectural layouts with graph neural networks, Comput. flexible industrial building structures integrating production planning, Buildings
Aided Civ. Inf. Eng. (2023), https://doi.org/10.1111/mice.13098. 12 (2) (2022) 162, https://doi.org/10.3390/buildings12020162.
[21] C.-T. Kupwiwat, K. Hayashi, M. Ohsaki, Deep deterministic policy gradient and [46] J.D. Rodriguez, M.E. Villafane, L. Piskorec, F.S. Caparrini, Generation of
graph attention network for geometry optimization of latticed shells, Appl. Intell. geometric interpolations of building types with deep variational autoencoders,
53 (2023) 19809–19826, https://doi.org/10.1007/s10489-023-04565-w. Design Sci. 6 (2020), https://doi.org/10.1017/dsj.2020.31.
[22] D. Shu, J. Cunningham, G. Stump, S.W. Miller, M.A. Yukish, T.W. Simpson, C. [47] J.-H. Jeong, H. Jo, Deep reinforcement learning for automated design of
S. Tucker, 3D design using generative adversarial networks and physics-based reinforced concrete structures, Comput. Aided Civ. Inf. Eng. 36 (12) (2021)
validation, J. Mech. Des. 142 (7) (2019), 071701, https://doi.org/10.1115/ 1508–1529, https://doi.org/10.1111/mice.12773.
1.4045419. [48] J.M. Mendel, Fuzzy logic systems for engineering: a tutorial, Proc. IEEE 83 (3)
[23] D. Sriram, M.L. Maher, S.J. Fenves, Knowledge-based expert systems in structural (1995) 345–377, https://doi.org/10.1109/5.364485.
design, Adv. Trends Struct. Dynamics (1985) 1–9, https://doi.org/10.1016/B978- [49] J.P. Amezquita-Sanchez, M. Valtierra-Rodriguez, M. Aldwaik, H. Adeli,
0-08-032789-1.50004-5. Neurocomputing in civil infrastructure, Scientia Iranica 23 (6) (2016)
[24] D.P. Kingma, M. Welling, Auto-encoding variational bayes, arXiv (2013), https:// 2417–2428, https://doi.org/10.24200/sci.2016.2301.
doi.org/10.48550/arXiv.1312.6114 preprint arXiv:1312.6114. [50] K. Hayashi, M. Ohsaki, Graph-based reinforcement learning for discrete cross-
[25] E. Kita, T. Toyoda, Structural design using cellular automata, Struct. Multidiscip. section optimization of planar steel frames, Adv. Eng. Inform. 51 (2022), 101512,
Optim. 19 (2000) 64–73, https://doi.org/10.1007/s001580050086. https://doi.org/10.1016/j.aei.2021.101512.
[26] F. Cascone, D. Faiella, V. Tomei, E. Mele, A structural grammar approach for the [51] K. Hayashi, M. Ohsaki, Reinforcement learning and graph embedding for binary
generative design of diagrid-like structures, Buildings 11 (3) (2021) 90, https:// truss topology optimization under stress and displacement constraints, Front.
doi.org/10.3390/buildings11030090. Built Environ. 6 (2020) 59, https://doi.org/10.3389/fbuil.2020.00059.
[27] F. Jiang, J. Ma, C.J. Webster, X. Li, V.J.L. Gan, Building layout generation using [52] K. Hayashi, M. Ohsaki, Reinforcement learning for optimum design of a plane
site-embedded GAN model, Autom. Constr. 151 (2023), 104888, https://doi.org/ frame under static loads, Eng. Comput. 37 (3) (2021) 1999–2011, https://doi.
10.1016/j.autcon.2023.104888. org/10.1007/s00366-019-00926-7.
[28] F.-A. Croitoru, V. Hondru, R.T. Ionescu, M. Shah, Diffusion models in vision: a [53] K. Shea, J. Cagan, Languages and semantics of grammatical discrete structures, AI
survey, IEEE Trans. Pattern Anal. Mach. Intell. (2023), https://doi.org/10.48550/ EDAM 13 (4) (1999) 241–251, https://doi.org/10.1017/S0890060499134012.
arXiv.2209.04747. [54] K. Shea, R. Aish, M. Gourtovaia, Towards integrated performance-driven
[29] G. Cheng, X. Zhou, J. Liu, L. Wang, Intelligent design method of high-rise shear generative design tools, Autom. Constr. 14 (2) (2005) 253–264, https://doi.org/
wall structures based on deep reinforcement learning, J. Build. Struct. 43 (09) 10.1016/j.autcon.2004.07.002.
(2022) 84–91, https://doi.org/10.14006/j.jzjgxb.2022.0067 (in Chinese). [55] K.H. Chang, C.Y. Cheng, Learning to simulate and design for structural
[30] G. Mirra, A. Pugnale, Comparison between human-defined and AI-generated engineering, in: International Conference on Machine Learning (ICML) vol. 119,
design spaces for the optimisation of shell structures, Structures 34 (2021) Electr Network, 2020, https://doi.org/10.48550/arXiv.2003.09103.
2950–2961, https://doi.org/10.1016/j.istruc.2021.09.058. [56] L. Wang, J. Liu, Y. Zeng, G. Cheng, H. Hu, J. Hu, X. Huang, Automated building
[31] H. Adeli, S. Kumar, Distributed genetic algorithm for structural optimization, layout generation using deep learning and graph algorithms, Autom. Constr. 154
J. Aerosp. Eng. 8 (3) (1995) 156–163, https://doi.org/10.1061/(ASCE)0893- (2023), 105036, https://doi.org/10.1016/j.autcon.2023.105036.
1321(1995)8:3(156). [57] M. Aalaei, M. Saadi, M. Rahbar, A. Ekhlassi, Architectural layout generation using
[32] H. Cao, C. Tan, Z. Gao, Y. Xu, G. Chen, P.-A. Heng, S.Z. Li, A survey on generative a graph-constrained conditional generative adversarial network (GAN), Autom.
diffusion model, arXiv (2022), https://doi.org/10.48550/arXiv.2209.02646 Constr. 155 (2023), 105053, https://doi.org/10.1016/j.autcon.2023.105053.
preprint arXiv:2209.02646. [58] M. Afzal, Y.H. Liu, J.C.P. Cheng, V.J.L. Gan, Reinforced concrete structural design
[33] H. Hofmeyer, J.M.D. Delgado, Coevolutionary and genetic algorithm based optimization: a critical review, J. Clean. Prod. 260 (2020), https://doi.org/
building spatial and structural design, Ai Edam-Artificial Intelligence for 10.1016/j.jclepro.2020.120623.
Engineering Design Analysis and Manufacturing 29 (4) (2015) 351–370, https:// [59] M. Aldwaik, H. Adeli, Advances in optimization of highrise building structures,
doi.org/10.1017/S0890060415000384. Struct. Multidiscip. Optim. 50 (6) (2014) 899–919, https://doi.org/10.1007/
[34] H. Omrany, A. Ghaffarianhoseini, R. Chang, A. Ghaffarianhoseini, F. Pour s00158-014-1148-1.
Rahimian, Applications of building information modelling in the early design [60] M. Huang, Z. Du, C. Liu, Y. Zheng, T. Cui, Y. Mei, X. Li, X. Zhang, X. Guo,
stage of high-rise buildings, Autom. Constr. 152 (2023), 104934, https://doi.org/ Problem-independent machine learning (PIML)-based topology optimization—a
10.1016/j.autcon.2023.104934. universal approach, Extreme Mech. Lett. 56 (2022), 101887, https://doi.org/
[35] H. Sun, H.V. Burton, H. Huang, Machine learning applications for building 10.1016/j.eml.2022.101887.
structural design and performance assessment: state-of-the-art review, J. Build. [61] M. Shahrouzi, M. Pashaei, Stochastic directional search: an efficient heuristic for
Eng. 33 (2021), 101816, https://doi.org/10.1016/j.jobe.2020.101816. structural optimization of building frames, Scientia Iranica 20 (4) (2013)
[36] H. Sun, L. Ma, Generative design by using exploration approaches of 1124–1132. https://scientiairanica.sharif.edu/article_1492_f3cea9b02cc96faaa
reinforcement learning in density-based structural topology optimization, Designs 2e26078828abcc7.pdf.
4 (2) (2020) 10, https://doi.org/10.3390/designs4020010. [62] M.L. Maher, Expert Systems for Civil Engineers: Technology and Application,
[37] H.L. Chi, X.Y. Wang, Y. Jiao, BIM-enabled structural design: impacts and future ASCE, 1987. ISBN:978–0–87262-617–1.
developments in structural modelling, analysis and optimisation processes, Arch.

17
W. Liao et al. Automation in Construction 157 (2024) 105187

[63] N. Nauata, K.-H. Chang, C.-Y. Cheng, C.-Y. Cheng, Y. Furukawa, House-GAN++: for early-stage building spatial design optimization, Autom. Constr. 124 (2021),
Generative adversarial layout refinement network towards intelligent https://doi.org/10.1016/j.autcon.2020.103522.
computational agent for professional architects, in: Proceedings of the IEEE/CVF [89] S. Chaillou, ArchiGAN: Artificial intelligence x architecture, in: P.F. Yuan, M. Xie,
Conference on Computer Vision and Pattern Recognition (CVPR), 2021, N. Leach, J. Yao, X. Wang (Eds.), Architectural Intelligence: Selected Papers from
pp. 13632–13641, https://doi.org/10.1109/CVPR46437.2021.01342. the 1st International Conference on Computational Design and Robotic
[64] N. Nauata, K.-H. Chang, C.-Y. Cheng, G. Mori, Y. Furukawa, House-GAN: Fabrication (CDRF 2019), Springer Nature Singapore, Singapore, 2020,
Relational Generative Adversarial Networks for Graph-constrained House Layout pp. 117–127, https://doi.org/10.1007/978-981-15-6568-7_8.
Generation, Springer International Publishing, Cham, 2020, pp. 162–177, [90] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput. 9 (8)
https://doi.org/10.1007/978-3-030-58452-8_10. (1997) 1735–1780, https://doi.org/10.1162/neco.1997.9.8.1735.
[65] N. Yüksel, H.R. Börklü, H.K. Sezer, O.E. Canyurt, Review of artificial intelligence [91] S. Krish, A practical generative design method, Comput. Aided Des. 43 (1) (2011)
applications in engineering design perspective, Eng. Appl. Artif. Intel. 118 (2023), 88–100, https://doi.org/10.1016/j.cad.2010.09.009.
105697, https://doi.org/10.1016/j.engappai.2022.105697. [92] S. Oh, Y. Jung, S. Kim, I. Lee, N. Kang, Deep generative design: integration of
[66] N.A. Kallioras, N.D. Lagaros, DzAIℕ: deep learning based generative design, topology optimization and generative models, J. Mech. Des. 141 (11) (2019),
Procedia Manuf. 44 (2020) 591–598, https://doi.org/10.1016/j. https://doi.org/10.1115/1.4044229.
promfg.2020.02.251. [93] S. Tafraout, N. Bourahla, Y. Bourahla, A. Mebarki, Automatic structural design of
[67] O. Ronneberger, P. Fischer, T. Brox, U-net: convolutional networks for biomedical RC wall-slab buildings using a genetic algorithm with application in BIM
image segmentation, in: Medical Image Computing and Computer-Assisted environment, Autom. Constr. 106 (2019), 102901, https://doi.org/10.1016/j.
Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, autcon.2019.102901.
October 5-9, 2015, Proceedings, Part III 18, Springer, 2015, pp. 234–241, https:// [94] S. Zhu, M. Ohsaki, K. Hayashi, X. Guo, Machine-specified ground structures for
doi.org/10.1007/978-3-319-24574-4_28. topology optimization of binary trusses using graph embedding policy network,
[68] P. Isola, J.-Y. Zhu, T. Zhou, A.A. Efros, Image-to-image translation with Adv. Eng. Softw. 159 (2021), 103032, https://doi.org/10.1016/j.
conditional adversarial networks, in: Proceedings of the IEEE Conference on advengsoft.2021.103032.
Computer Vision and Pattern Recognition, 2017, pp. 1125–1134, https://doi.org/ [95] S.K. Baduge, S. Thilakarathna, J.S. Perera, M. Arashpour, P. Sharafi, B. Teodosio,
10.48550/arXiv.1611.07004. A. Shringi, P. Mendis, Artificial intelligence and smart vision for building and
[69] P. Zakian, A. Kaveh, Seismic design optimization of engineering structures: a construction 4.0: machine and deep learning methods and applications, Autom.
comprehensive review, Acta Mech. 234 (4) (2023) 1305–1330, https://doi.org/ Constr. 141 (2022), https://doi.org/10.1016/j.autcon.2022.104440.
10.1007/s00707-022-03470-6. [96] T. Brown, B. Mann, N. Ryder, M. Subbiah, J.D. Kaplan, P. Dhariwal,
[70] P.J. Zhao, Y.F. Fei, Y.L. Huang, Y.T. Feng, W.J. Liao, X.Z. Lu, Design-condition- A. Neelakantan, P. Shyam, G. Sastry, A. Askell, Language models are few-shot
informed shear wall layout design based on graph neural networks, Adv. Eng. learners, Adv. Neural Inf. Proces. Syst. 33 (2020) 1877–1901, https://doi.org/
Inform. 58 (2023), 102190, https://doi.org/10.1016/j.aei.2023.102190. 10.48550/arXiv.2005.14165.
[71] P.J. Zhao, W.J. Liao, H.J. Xue, X.Z. Lu, Intelligent design method for beam and [97] T. Fischer, C.M. Herr, Teaching generative design, in: Proceedings of the 4th
slab of shear wall structure based on deep learning, J. Build. Eng. 57 (2022), Conference on Generative Art, Politechnico di Milano University Milan, 2001,
104838, https://doi.org/10.1016/j.jobe.2022.104838. pp. 147–160. https://www.researchgate.net/publication/30869860_Teachin
[72] P.J. Zhao, W.J. Liao, Y.L. Huang, X.Z. Lu, Intelligent beam layout design for frame g_Generative_Design.
structure based on graph neural networks, J. Build. Eng. 63 (2023), 105499, [98] T.-C. Wang, M.-Y. Liu, J.-Y. Zhu, A. Tao, J. Kautz, B. Catanzaro, High-resolution
https://doi.org/10.1016/j.jobe.2022.105499. image synthesis and semantic manipulation with conditional gans, Proc. IEEE
[73] P.J. Zhao, W.J. Liao, Y.L. Huang, X.Z. Lu, Intelligent design of shear wall layout Conf. Comput. Vis. Pattern Recognit. (2018) 8798–8807, https://doi.org/
based on attention-enhanced generative adversarial network, Eng. Struct. 274 10.1109/CVPR.2018.00917.
(2023), 115170, https://doi.org/10.1016/j.engstruct.2022.115170. [99] T.N. Kipf, M. Welling, Semi-supervised classification with graph convolutional
[74] P.J. Zhao, W.J. Liao, Y.L. Huang, X.Z. Lu, Intelligent design of shear wall layout networks, arXiv (2016), https://doi.org/10.48550/arXiv.1609.02907 preprint
based on graph neural networks, Adv. Eng. Inform. 55 (2023), 101886, https:// arXiv:1609.02907.
doi.org/10.1016/j.aei.2023.101886. [100] W. Hamilton, Z. Ying, J. Leskovec, Inductive representation learning on large
[75] P.N. Pizarro, L.M. Massone, F.R. Rojas, R.O. Ruiz, Use of convolutional networks graphs, Adv. Neural Inf. Process. Syst. 30 (2017), https://doi.org/10.48550/
in the conceptual structural design of shear wall buildings layout, Eng. Struct. 239 arXiv.1706.02216.
(2021), 112311, https://doi.org/10.1016/j.engstruct.2021.112311. [101] W. Huang, H. Zheng, Architectural drawings recognition and generation through
[76] P.N. Pizarro, L.M. Massone, Structural design of reinforced concrete buildings machine learning, in: Proceedings of the 38th Annual Conference of the
based on deep neural networks, Eng. Struct. 241 (2021), 112377, https://doi.org/ Association for Computer Aided Design in Architecture, Mexico City, Mexico,
10.1016/j.engstruct.2021.112377. 2018, pp. 18–20, https://doi.org/10.52842/conf.acadia.2018.156.
[77] P.N. Pizarro, N. Hitschfeld, I. Sipiran, J.M. Saavedra, Automatic floor plan [102] W. Ma, X. Wang, J. Wang, X. Xiang, J. Sun, Generative design in building
analysis and recognition, Autom. Constr. 140 (2022), 104348, https://doi.org/ information modelling (BIM): approaches and requirements, Sensors 21 (16)
10.1016/j.autcon.2022.104348. (2021) 5439, https://doi.org/10.3390/s21165439.
[78] R. Danhaive, C.T. Mueller, Design subspace learning: structural design space [103] W.J. Liao, X.Y. Wang, Y.F. Fei, Y.L. Huang, L.L. Xie, X.Z. Lu, Base-isolation design
exploration using performance-conditioned generative modeling, Autom. Constr. of shear wall structures using physics-rule-co-guided self-supervised generative
127 (2021), 103664, https://doi.org/10.1016/j.autcon.2021.103664. adversarial networks, Earthquake Engineering & Structural Dynamics 52 (11)
[79] R. Hu, Z. Huang, Y. Tang, O.V. Kaick, H. Zhang, H. Huang, Graph2Plan: learning (2023) 3281–3303, https://doi.org/10.1002/eqe.3862.
floorplan generation from layout graphs, ACM Trans. Graph. 39 (4) (2020) 118, [104] W.J. Liao, X.Z. Lu, Y.L. Huang, Z. Zheng, Y.Q. Lin, Automated structural design of
https://doi.org/10.48550/arXiv.2004.13204. shear wall residential buildings using generative adversarial networks, Autom.
[80] R. Kicinger, T. Arciszewski, K. De Jong, Parameterized versus generative Constr. 132 (2021), 103931, https://doi.org/10.1016/j.autcon.2021.103931.
representations in structural design: an empirical comparison, in: Genetic and [105] W.J. Liao, Y.L. Huang, Z. Zheng, X.Z. Lu, Intelligent generative structural design
Evolutionary Computation Conference, Washington, DC, 2005, pp. 2007–2014, method for shear wall building based on “fused-text-image-to-image” generative
https://doi.org/10.1145/1068009.1068344. adversarial networks, Expert Syst. Appl. 210 (2022), 118530, https://doi.org/
[81] R. Luo, Y. Wang, W. Xiao, X. Zhao, AlphaTruss: Monte Carlo tree search for 10.1016/j.eswa.2022.118530.
optimal truss layout design, Buildings 12 (5) (2022) 641, https://doi.org/ [106] W.X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, B. Zhang, J. Zhang,
10.3390/buildings12050641. Z. Dong, A survey of large language models, arXiv (2023), https://doi.org/
[82] R. Luo, Y. Wang, Z. Liu, W. Xiao, X. Zhao, A reinforcement learning method for 10.48550/arXiv.2303.18223 preprint arXiv:2303.18223.
layout design of planar and spatial trusses using kernel regression, Appl. Sci. 12 [107] X. Jia, R. Jiang, H. Qi, J. Liu, Z. Wu, A Multi-style Interior Floor Plan Design
(16) (2022) 8227, https://doi.org/10.3390/app12168227. Approach Based on Generative Adversarial Networks, Springer Nature Singapore,
[83] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High-resolution image Singapore, 2023, pp. 491–506, https://doi.org/10.1007/978-981-99-5844-3_36.
synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF [108] R. Luo Zhao, Structural shape grammars used in intelligent generation design of
Conference on Computer Vision and Pattern Recognition, 2022, discrete structures, in: 4th International Conference on Civil and Building
pp. 10684–10695, https://doi.org/10.48550/arXiv.2112.10752. Engineering Informatics, Sendai, Miyagi, Japan, 2019, pp. 325–329. https://
[84] R.E. Perez, K. Behdinan, Particle swarm approach for structural design www.researchgate.net/publication/337884833_Structural_Shape_Grammars_u
optimization, Comput. Struct. 85 (19–20) (2007) 1579–1588, https://doi.org/ sed_in_Intelligent_Generation_Design_of_Discrete_Structures.
10.1016/j.compstruc.2006.10.013. [109] X.Z. Lu, W.J. Liao, Y. Zhang, Y.L. Huang, Intelligent structural design of shear
[85] R.S. Sutton, A.G. Barto, Reinforcement Learning: An Introduction, MIT Press, wall residence using physics-enhanced generative adversarial networks,
2018. ISBN 0–262–19398-1. Earthquake Engineering & Structural Dynamics 51 (7) (2022) 1657–1676,
[86] S. Ampanavos, M. Nourbakhsh, C.-Y. Cheng, Structural Design Recommendations https://doi.org/10.1002/eqe.3632.
in the Early Design Phase using Machine Learning, Springer Singapore, [110] Y. Jia, J. Wang, W. Shou, M.R. Hosseini, Y. Bai, Graph neural networks for
Singapore, 2022, pp. 190–202, https://doi.org/10.48550/arXiv.2107.08567. construction applications, Autom. Constr. 154 (2023), 104984, https://doi.org/
[87] S. Boonstra, K. van der Blom, H. Hofmeyer, M.T.M. Emmerich, Conceptual 10.1016/j.autcon.2023.104984.
structural system layouts via design response grammars and evolutionary [111] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015)
algorithms, Autom. Constr. 116 (2020), https://doi.org/10.1016/j. 436–444, https://doi.org/10.1038/nature14539.
autcon.2019.103009. [112] Y. Zhang, C. Mueller, Shear wall layout optimization for conceptual design of tall
[88] S. Boonstra, K. van der Blom, H. Hofmeyer, M.T.M. Emmerich, Hybridization of buildings, Eng. Struct. 140 (2017) 225–240, https://doi.org/10.1016/j.
an evolutionary algorithm and simulations of co-evolutionary design processes engstruct.2017.02.059.

18
W. Liao et al. Automation in Construction 157 (2024) 105187

[113] Y.F. Fei, W.J. Liao, S. Zhang, P.F. Yin, B. Han, P.J. Zhao, X.Y. Chen, X.Z. Lu, [118] Y.T. Feng, Y.F. Fei, Y.Q. Lin, W.J. Liao, X.Z. Lu, Intelligent generative design for
Integrated schematic design method for shear wall structures: a practical shear wall cross-sectional size using rule-embedded generative adversarial
application of generative adversarial networks, Buildings 12 (9) (2022) 1295, network, J. Struct. Eng. 149 (11) (2023) 04023161, https://doi.org/10.1061/
https://doi.org/10.3390/buildings12091295. JSENDH.STENG-12206.
[114] Y.F. Fei, W.J. Liao, X.Z. Lu, H. Guan, Knowledge-enhanced graph neural networks [119] Z. He, T. Liang, X. Lai, M.X. Gao, X. Tu, Y. Lu, Vibration acceleration-integrated
for construction material quantity estimation of reinforced concrete buildings, parameterized aerodynamic shape optimization of super high-rise buildings with
Comput. Aided Civ. Inf. Eng. (2023), https://doi.org/10.1111/mice.13094. spiral configurations, Struct. Multidiscip. Optim. 65 (10) (2022), https://doi.org/
[115] Y.F. Fei, W.J. Liao, X.Z. Lu, T. Ertugrul, G. Hong, Semi-supervised learning 10.1007/s00158-022-03387-7.
method incorporating structural optimization for shear wall structure design [120] Z.J. Li, H. Adeli, Control methodologies for vibration control of smart civil and
using small and long-tailed datasets, J. Build. Eng. (2023), https://doi.org/ mechanical structures, Expert. Syst. 35 (6) (2018), https://doi.org/10.1111/
10.1016/j.jobe.2023.107873. exsy.12354.
[116] Y.F. Fei, W.J. Liao, Y.L. Huang, X.Z. Lu, Knowledge-enhanced generative [121] P.N. Pizarro, L.M. Massone, F.R. Rojas, Simplified shear wall building model for
adversarial networks for schematic design of framed tube structures, Autom. design optimization, J. Build. Eng. 76 (2023) 107368, https://doi.org/10.1016/j.
Constr. 144 (2022), 104619, https://doi.org/10.1016/j.autcon.2022.104619. jobe.2023.107368.
[117] Y.T. Feng, J.Y. Zhu, W.J. Liao, N. Zhao, Y.F. Fei, C.G. Li, P.J. Zhao, X.Z. Lu,
Construction bidding case study using generative intelligent structural design,
Build. Struct. (2023) (in Chinese).

19