Highlights in Science, Engineering and Technology CSIC 2023

Volume 85 (2024)

Utilizing Machine Learning Algorithms for Architectural Design

Katrina Sung
Pacific American School, Hsinchu 300, Taiwan, China
[email protected]

Abstract. With the development of digital technology, artificial intelligence has brought a paradigm
shift in numerous fields. Machine learning is used as a decision making tool to analyze large data
and produce output data with similar characteristics. This paper aims to explore and analyze the
influence of artificial intelligence on the architectural designing process. In the second section,
traditional methods of architectural design are observed. The third section explains the methodology
of machine learning and its current limitations. The fourth section investigates several real world
applications of machine learning in designing architecture, and the last section speculates on the
potential directions that the field of architectural design could take as machine learning continues to
advance.
Keywords: Machine Learning, architectural design.

1. Introduction
The field of architectural design holds an essential role in shaping the built environment,
harmonizing aesthetics, functionality, and sustainability. For decades, architects have wielded their
creativity to craft spaces that align with human experiences. Traditionally, architectural design has
been approached through manual drafting and physical models, where architects meticulously
transform their visions onto paper and materials. This craftsmanship, guided by experience and
artistic skills, has been the cornerstone of architectural design.
However, as technology develops, the investigation into Artificial Intelligence (AI) and its
diverse applications has brought a paradigm shift in their traditional approaches. Its indispensability
in modern times arises from its remarkable ability to address intricate problems that require
substantial time and effort. Functioning as a versatile tool, AI advances the process of integrating
information, analyzing data, and generating solutions to everyday problems. This technology has been
integrated with a variety of sectors, which includes areas such as healthcare and smart cities. Amid
this expansive landscape, one particularly promising application lies in the realm of architectural
design. Currently, about seven percent of the world’s labor force is in the construction industry.
However, it has fundamentally been one of the least technologically advanced industries due to the
complexity of designing progress. This juxtaposition creates an opportune arena for the convergence
of AI with architecture, a partnership that holds the potential to streamline design procedures and
significantly enhance overall efficiency.
Hence, this paper aims to closely examine how AI can influence architectural designs and how
it might transform the construction industry in the next decade. It evaluates both the traditional
methods and new machine learning algorithms used for architectural design and discusses the
potential directions that the field of architectural design could take as machine learning continues to
advance.

2. Traditional Methods
Some of the traditional methods such as manual drafting, physical models, and trial-and-error
methodologies have historically formed the bedrock of architectural creation.
Manual drafting involves the painstaking process of transforming conceptual ideas onto paper
using precision instruments. These skills often require extensive knowledge of the history of

26
Highlights in Science, Engineering and Technology CSIC 2023
Volume 85 (2024)

architecture as well as culture to achieve the desired outcome. This method, though emblematic of
craftsmanship, is fraught with challenges such as human error and time-intensive modifications. The
second method, physical models, provides tangible representations of architectural concepts.
Architects construct scaled-down models to showcase spatial relationships and aesthetics of buildings.
However, this method demands considerable time and resources, hindering the rapid exploration of
multiple design variations. The third approach, the trial-and-error method, allows architects to
produce an outcome that satisfies the consumer’s needs. However, the process is often resource-
consuming, which limits the exploration of design alternatives.
Eventually, the handcraft approach has evolved into a digital one. By the 1980s, the traditional
methods were replaced by the digital approach (Donaldson, 2023). Architects’ tools transitioned from
sheets of paper to software that can generate drawing outputs. This major shift has greatly impacted
the field and will hold as advanced technologies have replaced the simplicity of the pen. According
to Bill Kwon, vice president of IT at the architecture firm CallisonRTKL, “there are many
inefficiencies within the traditional architecture process in terms of design, delivery and simulation”
(Brodsky, 2020). He believes that AI will improve those problems through optimized methods that
can create efficiency.

3. Methodology
The term "Artificial Intelligence" was coined by John McCarthy in 1956. He characterized it as
"the discipline and engineering of creating intelligent machines" (McCarthy, 1990). It leverages
computers and algorithms to comprehend human intelligence and can thus be applied to tasks
involving problem-solving and decision-making. AI ought to possess the capability to execute the
design process and fulfill consumers' preferences. It can serve as a tool for clients to pinpoint the most
appropriate design that aligns with their desires. Imdat As, an architect specializing in digital design
and holding the position of assistant professor of architecture, elucidates that deep-learning machines
and AI have the potential to alter the longstanding patterns in architectural design (Imdat, 2018). AI
could potentially expand upon these patterns, encompassing functional and programmatic
considerations while incorporating additional variables such as geographical and climatic factors that
influence the built environment. Through AI, machines can scrutinize the optimal patterns for a given
problem and substantiate their findings. The design process stands to significantly benefit from
heightened intelligence and efficiency. Early research relating to the use of machine learning in
architectural design focused on design generation, shape recognition, and space organization (Tamek,
2018).
3.1 Machine Learning
Machine Learning is a multidisciplinary field that focuses on the development of algorithms and
statistical models that enable computer systems to improve their performance through learning from
data, without being explicitly programmed. This paper will focus on neural networks.
A neural network is a computational system composed of algorithms that aims to uncover hidden
patterns within a dataset by imitating the way the human brain functions. Whether made up of
biological neurons or artificial counterparts, these networks can adjust to varying inputs, allowing
them to produce optimal outcomes without the need for a complete redesign of their output rules.

27
Highlights in Science, Engineering and Technology CSIC 2023
Volume 85 (2024)

Figure 1: Neural Networks

3.2 Types of machine learning used in architectural design
3.2.1 Generative adversarial networks (GAN)
Generative adversarial networks (GAN) is a model framework that can learn from input data and
generate outputs with similar characteristics. There are two components in the model, the Generator
and the Discriminator, which allow the model to operate in a feedback loop to refine the model’s
ability to generate outputs. While the Discriminator is used to recognize images from data, the
Generator can create images that are similar to the ones in the data. In response, the Discriminator
then provides feedback for the Generator to improve on. As this process continues, a GAN is trained
to create synthetic images (Goyal, 2019).
GANs are useful in generating new architectural designs and adapting existing ones to different
styles, offering architects creative inspiration and flexibility. They contribute to interior design by
generating layouts and decor suggestions and aid in parametric design optimization. Moreover, GANs
create high-quality 3D models and immersive VR/AR experiences, enhancing design
communication.

28
Highlights in Science, Engineering and Technology CSIC 2023
Volume 85 (2024)

Fig.2 Generative Adversarial Network Model Architecture

Source: Mohamed, et al. (2022)
3.2.2 Convolutional neural network (CNN)
Convolutional Neural Networks (CNNs) are integral to modern architectural design processes.. It
is a multilayer neural network that’s often used in image-processing applications. Essentially, CNN
is made up of multiple layers of feature extraction and classification processes. When inputting an
image, the image is divided into areas that feed into a convolutional layer, which extracts features
from it. Then, the pooling process reduces the dimension of the extracted features. The next stage
involves another round of convolution and pooling, which then connects to a fully connected
multilayer perceptron. This network's ultimate output layer consists of nodes that recognize specific
characteristics within the image. Training the network is accomplished through the iterative process
of back-propagation (IBM Developer, 2023).
CNN enables image recognition and classification, aiding architects in categorizing styles and
features of existing architectures. It also offers design inspiration by analyzing architectural images
and suggesting elements aligned with project goals. In design optimization, they help fine-tune
layouts and materials for enhanced sustainability and functionality. CNNs recognize materials and
textures, ensuring suitable choices, and can even assess environmental impact. They also contribute
to safety compliance by identifying hazards in designs.
3.3 Limits of machine learning in its current state
Architectural processes using AI need constant supervision. Hence, one of the limitations is the
possible occurrence of human error in the process, which can be reduced by hiring experts to
supervise the work. Some of the drawbacks of using AI in architecture include dependence on
technology, potential job loss, as well as privacy and security risks (Asri).

29
Highlights in Science, Engineering and Technology CSIC 2023
Volume 85 (2024)

4. Evaluation
4.1 Architectural AI
With the growing power of artificial intelligence, many corporations have been conducting
research on AI’s potential to transform urban spaces. Ravi Bedi, Technology Innovation Principal
Director at Accenture, highlights two fundamental changes that make AI important in the future,
which are the growing trend of digitization and the reduction in the cost of processing data (Dhawan,
2020). During the designing process, architects review past designs and the data prepared throughout
the making of the building. AI will analyze data and give insights which the architects will be able to
do testing in a faster way. There are several ways AI can shape future architecture. For example, it
uses geometry programming with complex algorithms that allow architects to change parameters to
create innovative designs (Reddy, 2020). Moreover, it can save an architect’s time and energy to plan
by analyzing the whole data in milliseconds and creating foundation models so that the architects can
add details to it (Kai, 2019).
According to statistics, “About seven percent of the world’s labor force is in the construction
industry”, however, “it has traditionally been one of the least technologically advanced industries”
(Kai, 2019). Nevertheless, experts believe that there is great potential for the integration of AI in the
architecture industry as it could reduce the building and designing costs by up to twenty percent.
"There are many inefficiencies within the architecture process in the realms of design, delivery, and
simulation," Bill Kwon, vice president of IT and digital transformation at the architecture firm
CallisonRTKL, said in an email interview. "AI will inevitably wean out the waste through optimized
and predictive modeling, create efficiencies, and reduce errors through automation and continuous
learning” (Brodsky, 2020).
Research has been conducted using a deep neural network (DNN), which is a deep learning
application that “extracts design into essential building blocks - based on functional performance
criteria - and recombines them into new designs” (As, 2018). DNNs are not bound by rules or original
algorithms but can learn patterns through training, which is different from conventional approaches.
In architecture, a project goes through several phases, including conceptual design, schematic design,
design development, construction documents, procurement, construction administration, and
operations. Deep learning can help shorten the conceptual design process. It can use graphs to
generate conceptual designs. This tool can evaluate design models, input them into building blocks,
and yield them into new design compositions.
4.2 Generative design
Stanislas Chaillou, a designer, worked on a project at Harvard aiming to apply machine learning
in architectural practice. He focused on the area of AI-generated floor plans and developed training
models aligning to specific architectural styles: Baroque, Row House, Victorian Suburban House, &
Manhattan Unit. The study of the style is essential as it demonstrates how a particular style can “define
mechanics and space” and “control the internal organization of the plan” (Chaillou, 2019). Next, he
used a framework, Pix2Pix, which is a standard generative adversarial neural network (GAN) model.
There are two components in the model, the Generator and the Discriminator, which allow the model
to operate in a feedback loop to refine the model’s ability to generate floor plans. While the
Discriminator is used to recognize images from data, the Generator can create images that are similar
to the ones in the data. In response, the Discriminator then provides feedback for the Generator to
improve on. As this process continues, a GAN is trained to create synthetic images.
Chaillou adopted the model from this project in a large-scale housing development project located
in Manhattan’s Lower East Side. This project explores the potential of AI in floor plan design: firstly,
it can aid architects in creating cohesive room arrangements and furnishings; secondly, it can
reconstruct all apartment units into a preliminary floor plan; and thirdly, transform floor plans from
one style to another.

30
Highlights in Science, Engineering and Technology CSIC 2023
Volume 85 (2024)

5. Conclusion and Future Work

As AI technology advances, the architectural design field will continue to be affected. Emerging
technologies, such as real-time simulation, augmented reality, and integrative design, hold the
promise of reshaping the future of architectural practice. Fueled by advanced machine learning
algorithms, real-time simulation can provide architects with dynamic, interactive models that respond
to real-world conditions, facilitating more accurate design decisions. Augmented reality, seamlessly
integrated with architectural workflows, could offer architects immersive experiences, allowing them
to visualize and modify designs on-site, enhancing precision and client engagement. Integrative
design, propelled by machine learning's data-driven insights, encourages cross-disciplinary
collaboration, ensuring that architecture evolves in harmony with other domains like sustainability,
engineering, and urban planning. These technological advancements, coupled with the creative
ingenuity of architects, herald an era where design becomes more efficient, adaptable, sustainable,
and attuned to the evolving needs of society.
The direction of AI in architectural development is not certain, but multiple researches have been
conducted to collect the opinions of designers and researchers with expertise in artificial intelligence
regarding the future development of architectural design. Imdat As, an assistant professor of
architecture with expertise in digital design, believes that AI can drastically improve our built
environment since architects have more time to “reach a much wider audience and provide access to
quality design to a broader portion of the public” (As, 2018). However, some interviewees believe
that AI still lacks accessibility in architecture because of the lack of data and technology. Nonetheless,
experts agree that artificial intelligence “will be highly useful for professional practice over the next
ten years” (Gallo, 2020). To facilitate this transformation, AI integration into the workflows of
modern design firms is imperative, and architects must acquire the necessary skills to effectively
employ these tools, ensuring they remain at the forefront of their field.

References
[1] Asri, Youne. AI in Architecture: Exploring the Benefits, Drawbacks, and Innovative Differences.
https://bootcamp.uxdesign.cc/ai-in-architecture-exploring-the-benefits-drawbacks-and-innovative-
differences-b1a5a719a2bc.
[2] Brodsky, S. (2020, December 9). How AI is Changing Architecture. Lifewire. Retrieved April 13, 2022,
from https://www.lifewire.com/how-ai-is-changing-architecture-5090676
[3] Chaillou, Stanislas. “AI Creates Generative Floor Plans and Styles with Machine Learning at Harvard |
ArchDaily.” ArchDaily, ArchDaily, 5 June 2019, https://www.archdaily.com/918471/ai-creates-
generative-floor-plans-and-styles-with-machine-learning-at-harvard.
[4] Dhawan, Vidushi. “Artificial Intelligence Will Have a Direct Impact on Future Urban Planning –
OpenGov Asia.” OpenGov Asia, 2 Mar. 2020, https://opengovasia.com/artificial-intelligence-will-have-
a-direct-impact-on-future-urban-planning/.
[5] Donaldson, James E. With or Without AI; The Architect’s Methodology. 20 June 2023,
https://www.linkedin.com/pulse/without-ai-architects-methodology-james-jed-donaldson/.
[6] Dr. Kai Oberste-Ufer. “7 Ways Artificial Intelligence Is Revolutionizing Architecture.” EN - Dormakaba
Blog, https://www.facebook.com/dormakabaGLOBAL/, 2 Sept. 2019, https://blog.dormakaba.com/7-
ways-artificial-intelligence-is-revolutionizing-architecture/.
[7] Eltay, Mohamed, et al. “Generative Adversarial Network Based Adaptive Data Augmentation for
Handwritten Arabic Text Recognition.” PeerJ Computer Science, PeerJ, Jan. 2022, p. e861. Crossref,
doi:10.7717/peerj-cs.861.
[8] Gallo, Giuseppe, and wirz. “(PDF) The Role of Artificial Intelligence in Architectural Design:
Conversation with Designers and Researchers | Giuseppe Gallo and Fulvio Wirz - Academia.Edu.”
Academia.Edu - Share Research, 1 Jan. 2020,
https://www.academia.edu/44902106/The_role_of_Artificial_Intelligence_in_architectural_design_conv
ersation_with_designers_and_researchers.

31
Highlights in Science, Engineering and Technology CSIC 2023
Volume 85 (2024)

[9] Goyal, Shweta. “GANs — A Brief Introduction to Generative Adversarial Networks | by Shweta Goyal |
Analytics Vidhya | Medium.” Medium, Analytics Vidhya, 2 June 2019, https://medium.com/analytics-
vidhya/gans-a-brief-introduction-to-generative-adversarial-networks-f06216c7200e.
[10] “IBM Developer.” IBM Developer, https://developer.ibm.com/articles/cc-machine-learning-deep-
learning-architectures/. Accessed 22 Sep. 2023.
[11] Mahesh, Batta. “Machine Learning Algorithms - A Review.” International Journal of Science and
Research , Jan. 2020.
[12] McCarthy, John. Formalizing Common Sense. Intellect Books, 1990.
[13] Reddy, Vinod. “How Is Artificial Intelligence (AI) Changing the Future of Architecture?” AiThority,
https://www.facebook.com/AiThority, 16 Jan. 2020, https://aithority.com/guest-authors/how-is-artificial-
intelligence-ai-changing-the-future-of-architecture/.
[14] Tamke, Martin, et al. “Machine Learning for Architectural Design: Practices and Infrastructure.”
International Journal of Architectural Computing, June 2018

32