arXiv copy T.

Kapsalis (2024)

CADgpt: Harnessing Natural Language Processing for 3D Modelling

to Enhance Computer-Aided Design Workflows
Authors
Timo Kapsalis[a] – t.kapsalis@derby.ac.uk
[a]
: University of Derby, UK

Abstract
This paper introduces CADgpt, an innova�ve plugin integra�ng Natural Language Processing (NLP)
with Rhino3D for enhancing 3D modelling in computer-aided design (CAD) environments. Leveraging
OpenAI's GPT-4, CADgpt simplifies the CAD interface, enabling users, par�cularly beginners, to
perform complex 3D modelling tasks through intui�ve natural language commands. This approach
significantly reduces the learning curve associated with tradi�onal CAD so�ware, fostering a more
inclusive and engaging educa�onal environment. The paper discusses CADgpt's technical
architecture, including its integra�on within Rhino3D and the adapta�on of GPT-4 capabili�es for
CAD tasks. It presents case studies demonstra�ng CADgpt's efficacy in various design scenarios,
highligh�ng its poten�al to democra�se design educa�on by making sophis�cated design tools
accessible to a broader range of students. The discussion further explores CADgpt's implica�ons for
pedagogy and curriculum development, emphasising its role in enhancing crea�ve explora�on and
conceptual thinking in design educa�on.

Keywords: Natural Language Processing, Computer-Aided Design, 3D Modelling, Design Automa�on,

Design Educa�on, Architectural Educa�on

Introduction
In the evolving narrative of Computer-Aided Design (CAD), the interface between
human creativity and computational precision has been a subject of continuous
refinement [1,2]. The trajectory from command-line to graphical user interfaces in
CAD software marked a significant leap in usability and accessibility [3]. Yet, the
integration of Artificial Intelligence (AI), and specifically natural language processing
(NLP), promises a further democratisation of design tools, enabling a more intuitive
interaction that aligns with human cognitive processes [4,5]. We introduce CADgpt, a
groundbreaking plugin for Rhino3D that leverages OpenAI's GPT-4 model, to extend
the frontier of design education and practice.
The development of CADgpt is underpinned by a critical recognition of the
challenges faced by nascent design students when confronted with the complexity of
traditional CAD interfaces [6]. Previous evidence from educational settings reveals a
persistent barrier: the steep learning curve that can stifle the creative exploration of
students new to 3D modelling environments [6,7]. CADgpt aims to dismantle this
barrier by allowing users to employ natural language to execute complex modelling
tasks, thereby facilitating a more inclusive and engaging learning experience.
The purpose of CADgpt transcends mere technical innovation; it is a pedagogical
tool fashioned to integrate seamlessly into the design curriculum. By enabling

1
arXiv copy T. Kapsalis (2024)

command execution through conversational language, CADgpt aligns with cognitive

learning theories which posit that language is a critical mediator in the construction of
knowledge. This alignment suggests that by reducing the technical overhead of CAD
software, students can focus more on the creative aspects of design, fostering a
more profound educational experience.
The potential impact of CADgpt on the field of design education is profound. By
abstracting the technical complexities of 3D modelling into intuitive language,
CADgpt stands to broaden the accessibility of design education, enhance the
efficiency of learning, and enable a wider demographic of students to engage with
advanced design tools. This shift not only empowers students of diverse
backgrounds and abilities but also invites a reassessment of the role of AI in the
creative process, suggesting new paradigms of interaction between designers and
their tools.
In the following sections, this paper will detail the technical architecture of CADgpt,
illustrate its application through examples from educational settings, and discuss the
broader implications for design educators, students, and the future of design
education in an era increasingly characterised by the symbiosis of human creativity
and computational intelligence.
Technical description
CADgpt represents an interactive integration of OpenAI's GPT-4 model [8] with
Rhino3D CAD environment [9]. This section delineates the technical constitution of
CADgpt, elucidating the operational mechanics that empower users to engage with
Rhino3D through the medium of natural language.
CADgpt is architected as a plugin that functions within the Rhino3D software
ecosystem. It serves as an intermediary, translating natural language inputs into the
Grasshopper scripting and command framework. This translation is facilitated
through a custom API that leverages the GPT-4 model's advanced NLP capabilities.
The API acts as a conduit, parsing user inputs, interpreting their design intentions,
and generating corresponding RhinoScript or Grasshopper definitions that Rhino3D
can execute.
At the heart of CADgpt is the GPT-4 AI model, a state-of-the-art language processor
renowned for its ability to understand and generate human-like text [8]. The engine
processes input commands in natural language, comprehends the context and
intent, and translates these into a structured command set that is compatible with
CAD operations. This process includes an understanding of geometric concepts,
design terminology, and procedural instructions, all within the vernacular of design
automation.
A typical interaction with CADgpt begins with the user articulating a design task in
natural language. This input is processed by GPT-4, which generates a script that
includes necessary variables, functions, and logic. The output script is then passed

2
arXiv copy T. Kapsalis (2024)

to Rhino3D's command processor. For example, a user may input "create a 3D

model of a ten-story building with a glass facade", which CADgpt will interpret and
translate into a series of Rhino3D actions, creating the geometric model as specified.
Figure 1 presents a workflow with basic operations of the CADgpt tool.

Figure 1: The CADgpt operational workflow

CADgpt employs custom algorithms (e.g., Natural Language Understanding and

Geometric Modelling algorithms) to refine the output from the GPT-4 model, ensuring
the generated commands are optimised for Rhino3D's environment. These
algorithms address the specificities of CAD tasks, such as the precision required in
geometric calculations and the hierarchical structuring of design elements. The
optimisation process also includes error checking and feedback loops, which allow
CADgpt to learn from interactions and improve over time.
The CADgpt plugin is designed with a minimalist interface to maintain focus on the
Rhino3D environment while providing clear and concise feedback to the user. Visual
cues and command confirmation are built into the system to guide users through the
design process and ensure clarity in command execution.
In the construction of CADgpt, careful attention has been paid to the ethical
implications of AI integration in design tools. All interactions with the GPT-4 model
are conducted in accordance with OpenAI's use-case policies, ensuring data privacy
and security [8]. Additionally, the plugin includes safeguards to prevent the
generation of designs that could be ethically problematic or unsafe.
The technical architecture of CADgpt is a testament to the harmonious potential
between human-centric design processes and AI efficiency. By enabling natural
language as a means to command and control within a CAD environment, CADgpt
stands as a significant advance in the field of design automation, poised to redefine
the interface between designers and digital creation tools.

3
arXiv copy T. Kapsalis (2024)

Early examples of CADgpt implementation

The practical applications of CADgpt are best illustrated through case studies that
highlight its functionality and user experience. Here we present three such examples,
showcasing the tool's use in varying contexts. All examples refer to student exercises
within the Architectural Technology program at the University of Derby.
Example 1: geometric interplay
The attached image demonstrates a fundamental operation where a student used
CADgpt to create a complex geometric model. The task was to intersect a
100x100x30 cm box with a sphere of 30 cm radius, positioned at a random edge,
and then execute a union of the two shapes within Rhino3D. The student simply
input the natural language command: "Create a 100x100x30 cm box, which is
intersected by a sphere of 30 cm radius at a random edge. Bake their union on
Rhino". CADgpt interpreted this command, generated the corresponding geometric
shapes, and performed the Boolean union operation—all without the student needing
to manually input any traditional Rhino3D commands or navigate complex toolbars
(Figure 2).

Figure 2: Using CADgpt for simple geometric modelling, an example.

4
arXiv copy T. Kapsalis (2024)

Example 2: architectural prototyping

In an architectural design workshop, students were tasked with conceptualising a
pavilion with organic forms, taking inspiration from precedent projects. Using
CADgpt, one student described their vision: "Design a pavilion with a hyperbolic
canopy, inspired by the Candela structures". CADgpt processed this description and
provided a step-by-step Grasshopper definition, which iteratively refined the
pavilion's form based on the student's further natural language feedback, allowing for
a rapid prototyping process that typically would require advanced Rhino3D
proficiency (Figure 3).

Figure 3: Using CADgpt for early-stage architectural prototyping, an example.

5
arXiv copy T. Kapsalis (2024)

Example 3: urban design simulation

During a course on sustainable urban design, students used CADgpt to simulate
sunlight patterns across a planned residential area. A student inputted: "Generate a
grid of buildings 15 meters high, spaced 20 meters apart, and simulate the sunlight
paths during the UK summer solstice". CADgpt recognised the command, created
the grid of buildings, and utilised Rhino3D's sun path simulation tools to display the
sunlight patterns (Figure 4). This allowed the student to conduct shade studies to
assess the potential for passive solar heating and make informed decisions about
building placement and design for energy efficiency.

Figure 4: Using CADgpt for urban design tasks, an example.

6
arXiv copy T. Kapsalis (2024)

These examples underscore the versatility of CADgpt in translating natural language

into complex CAD operations, enabling students to engage with advanced design
scenarios without the prerequisite of in-depth CAD knowledge. Each case study
reflects the tool's potential to expand the boundaries of design education, making
sophisticated design and analysis tasks accessible to a broader range of students.
Discussion
As CADgpt remains in a work-in-progress and trial stage, its introduction into the
design curriculum is a pioneering step that necessitates both cautious optimism and
critical reflection. This emergent tool offers an opportunity to fundamentally reassess
the pedagogical strategies employed in design education, with broad implications for
educators, students, and the wider field of design automation research.
CADgpt embodies the pedagogical principle that tool transparency can enhance
learning outcomes. By reducing the cognitive load associated with learning complex
software commands, educators can refocus their pedagogy on design thinking and
problem-solving [10,11]. CADgpt aligns with the constructivist approach to education,
which emphasises the importance of active, contextualised learning processes
where students construct new knowledge based on their experiences [12,13].
The implementation of CADgpt necessitates a revaluation of design curriculum
structures. Courses traditionally structured around the gradual introduction of
software functionality may need to be adapted to integrate AI tools like CADgpt,
which change the way students interact with CAD software. This integration also
opens the door to a more interdisciplinary curriculum that can include students from
various backgrounds, promoting diversity within the field of design.
One of the most compelling contributions of CADgpt is its potential to liberate student
creativity from the technical constraints of traditional CAD learning [14,15]. By
simplifying interaction with CAD software through natural language processing,
students can allocate more time and cognitive resources to conceptual development
and creative experimentation. As early results indicated, CADgpt facilitates a shift
away from the intricacies of command syntax and workflow memorisation, allowing
students to engage with the essence of design itself - the realisation of innovative
concepts and solutions.
For design students, especially those in the nascent stages of their education,
CADgpt offers a streamlined path to developing CAD proficiency. By interacting with
the software through natural language, students can engage with 3D modelling
without the intimidation of intricate toolsets [16]. It is therefore possible that this
accessibility can accelerate the acquisition of spatial reasoning and design skills,
which are fundamental to the architectural and design disciplines.
By lowering language and technical barriers, CADgpt contributes significantly to
inclusive education and adaptive learning. By providing a platform that responds to
natural language commands, CADgpt inherently adapts to the diverse cognitive and

7
arXiv copy T. Kapsalis (2024)

linguistic styles of its users. This adaptability ensures that students can engage with
the tool in a manner that aligns with their personal learning preferences, thus
supporting a more personalised educational journey. The adaptive nature of CADgpt
is particularly salient in its potential to cater to students with disabilities or those who
might otherwise find the technical rigor of traditional CAD interfaces a barrier. The
tool's capacity to learn and adjust to individual student interactions embodies the
principles of adaptive learning—where technology is used to tailor educational
experiences to each student's needs and abilities [17,18]. As such, the tool has the
potential to democratise design education, making it more accessible to a wider
range of learning styles and abilities.
CADgpt can serve as a spark for research into the intersection of AI and design. It
provides a practical example of how AI can be integrated into design tools, inviting
further investigation into how such integration can evolve. Research can explore the
implications of AI-assisted design on creativity, the development of new design
methodologies, and the future role of designers in an AI-augmented design process.
Looking beyond education to professional practice, CADgpt suggests a future where
designers may collaborate more closely with AI, utilising natural language to
communicate with CAD systems. This could lead to more intuitive design processes
and potentially change the role of designers, who may increasingly become
orchestrators of AI-assisted design rather than sole creators [19-21].
As CADgpt evolves, its development will be directed towards refining its ability to
understand and execute increasingly complex and abstract design requests. This
progression will entail training the AI with diverse datasets that capture the breadth of
design terminology and conceptual descriptions. The aim is to develop a tool that not
only understands detailed commands but can also engage with the ambiguous and
often non-linear nature of creative design processes. Further development of
CADgpt will focus on expanding its capabilities to interpret more complex design
tasks and integrating with additional specialised plugins, enhancing its versatility.
CADgpt represents a transformative tool that has the potential to reshape the
landscape of design education and practice. By enabling natural language
interaction with CAD software, it simplifies the learning process, fosters inclusivity,
and opens up new avenues for research and practice in design automation. As the
tool evolves and its usage becomes more widespread, it is likely to have a lasting
impact on how we teach, learn, and practice design in an increasingly digital world.

8
arXiv copy T. Kapsalis (2024)

References
[1] Samuelle, B., Jane, L.E. (2023). Exploring the Horizon of Computa�on for Crea�vity. Crossroads,
doi: htps://10.1145/3596921

[2] Ji, Y., Lee, K., Na, K. (2022). A Study on U�lisa�on of Computa�onal Thinking for Implemen�ng
Crea�vity. Design research, doi: htps://10.46248/kidrs.2022.4.388.

[3] Wenshi, S., Ketai, H. (2011). Design of interac�on of CAD pla�orm. In Proceedings of the 6th IEEE
Conference on Industrial Electronics and Applica�ons. Beijing, China, 2011, pp. 1482-1487, doi:
htps://10.1109/ICIEA.2011.5975824.

[4] Tse-Hao, L. (2023). Use of Language to Generate Architectural Scenery with AI-Powered Tools.
Communica�ons in Computer and Informa�on Science, doi: htps://10.1007/978-3-031-37189-9_6.

[5] Tholander, J., Jonsson, M. (2023). Design Idea�on with AI - Sketching, Thinking and Talking with
Genera�ve Machine Learning Models. In Proceedings of the 2023 ACM Designing Interac�ve Systems
Conference (DIS '23), Associa�on for Compu�ng Machinery, New York, NY, USA, doi:
htps://10.1145/3563657.3596014.

[6] Arteaga-Juárez, P., Dreifuss-Serrano, C., Herrera, P.C. (2022). Undergraduate percep�ons of
remote teaching in Computer-Aided Design courses. In Proceedings of the IEEE Learning with MOOCS
(LWMOOCS), doi: htps://10.1109/lwmoocs53067.2022.9927920.

[7] Beasley, Z., Piegl, L.A., Rosen, P. (2018). Ten challenges in CAD cyber educa�on. Computer-Aided
Design and Applica�ons, doi: htps://10.1080/16864360.2017.1397893.

[8] OpenAI. (2023). GPT-4 Technical Report. arXiv:2303.08774,

htps://doi.org/10.48550/arXiv.2303.08774.

[9] McNeel & Associates. (2023). Rhino3D, Version 8.1. Available at: htps://www.rhino3d.com/.

[10] Pruthviraj, C., Chavan, A.K., Pa�l, A.K., Chate, G.R., Malik, V. (2023). Design Thinking with ICT
Tools: An Approach to Enhance Engagement in Design Problem and Prac�cal Exposure. Journal of
Engineering Educa�on Transforma�ons, doi: htps://10.16920/jeet/2023/v36is2/23007.

[11] Lee, C.B., Skillen, M., Williams, C. (2023). Designing Problem-Solving for Meaningful Learning: A
Discussion of Asia-Pacific Research. In Lee, W.O., Brown, P., Goodwin, A.L., Green, A. (Eds.),
Interna�onal Handbook on Educa�on Development in Asia-Pacific. Springer, Singapore,
htps://doi.org/10.1007/978-981-16-2327-1_51-1.

[12] Mishra, N.R. (2023). Construc�vist Approach to Learning: An Analysis of Pedagogical Models of
Social Construc�vist Learning Theory. Journal of Research and Development, doi:
htps://10.3126/jrdn.v6i01.55227.

[13] Fiaz, N., Eddy, C.M. (2020). Construc�vist educa�on in an age of accountability. Journal of
Educa�on Policy, doi: htps://10.1080/02680939.2019.1619248.

[14] Parampalli, A., Viswanathan, V. (2019). A Study on the Role of Computer-aided Design in Design
Crea�vity and Educa�on. Engineering Design Graphics Journal.

[15] Arrighi, P.-A., Le Masson, P., Weil, B. (2015). Addressing Constraints Crea�vely: How New Design
So�ware Helps Solve the Dilemma of Originality and Feasibility. Crea�vity and Innova�on
Management, doi: htps://10.1111/CAIM.12082.

9
arXiv copy T. Kapsalis (2024)

[16] Fougères, A.-J., Ostrosi, E. (2020). Intelligent requirements engineering from natural language
and their chaining toward CAD models. arXiv: Ar�ficial Intelligence, doi:
htps://arxiv.org/abs/2007.07825.

[17] Dziuban, C.D., Howlin, C., Moskal, P., Johnson, C., Parker, L., Campbell, M. (2018). Adap�ve
Learning: A Stabilizing Influence Across Disciplines and Universi�es. Online Learning Consor�um, doi:
htps://10.24059/OLJ.V22I3.1465.

[18] Capuano, N., Caballé, S. (2020). Adap�ve Learning Technologies. AI Magazine, doi:
htps://10.1609/AIMAG.V41I2.5317.

[19] Chiou, L.-Y., Hung, P.-K., Liang, R.-H., Wang, C.-T. (2023). Designing with AI: An Explora�on of Co-
Idea�on with Image Generators. In Proceedings of the 2023 ACM Designing Interac�ve Systems
Conference (DIS '23), Associa�on for Compu�ng Machinery, New York, NY, USA,
htps://doi.org/10.1145/3563657.3596001.

[20] Figoli, F.A., Rampino, L., Ma�oli, F. (2022). AI in design idea development: A workshop on
crea�vity and human-AI collabora�on. In Lockton, D., Lenzi, S., Hekkert, P., Oak, A., Sádaba, J., Lloyd,
P. (Eds.), DRS2022: Bilbao, 25 June - 3 July, Bilbao, Spain, htps://doi.org/10.21606/drs.2022.414.

[21] Gmeiner, F., Yang, H., Yao, L., Holstein, K., Martelaro, N. (2023). Exploring Challenges and
Opportuni�es to Support Designers in Learning to Co-create with AI-based Manufacturing Design
Tools. In Proceedings of the 2023 CHI Conference on Human Factors in Compu�ng Systems (CHI '23),
Associa�on for Compu�ng Machinery, New York, NY, USA, Ar�cle 226, 1–20,
htps://doi.org/10.1145/3544548.3580999.

10