Experiments on Generative AI-Powered Parametric Modeling and BIM for

Architectural Design

a, * a a, *
Jaechang Ko , John Ajibefun , Wei Yan
a
Department of Architecture, Texas A&M University, College Station, 77840, TX, United States

Abstract

This paper introduces a new architectural design framework that utilizes generative AI tools
including ChatGPT and Veras with parametric modeling and Building Information Modeling
(BIM) to enhance the design process. The study experiments with the potential of ChatGPT and
generative AI in 3D architectural design, extending beyond its use in text and 2D image
generation. The proposed framework promotes collaboration between architects and AI,
facilitating a quick exploration of design ideas and producing context-sensitive, creative design
generation. By integrating ChatGPT for scripting and Veras for generating design ideas with
widely used parametric modeling and BIM tools, the framework provides architects with an
intuitive and powerful method to convey design intent, leading to more efficient, creative, and
collaborative design processes.

Keywords: Generative AI, ChatGPT, Generative Design, Design Automation, Computational

Design

1. Introduction
The architectural design process is complex and continuously evolving, utilizing various tools
and methodologies to support architects in modeling, visualizing, and analyzing their designs
(Cross 1982). With the rapid advancement of technology, artificial intelligence (AI) and machine
learning (ML) have been integrated into the design process, presenting new opportunities and
challenges for architects and designers. However, the potential for AI, particularly language
models like ChatGPT - a conversational AI model developed by OpenAI (Radford et al. 2021)-
to transform the architectural design process has yet to be fully explored. This paper presents a
new framework for architectural design that uses ChatGPT and AI-based ideation and
visualization tools, Veras (“VERAS” 2023), to make the design process easier and create 3D
geometric models, parametric models, and Building Information Models using natural language
input.
The proposed framework combines ChatGPT and Veras to generate and explore design ideas
rapidly. Using natural language input, architects can communicate their design intentions more
intuitively, allowing quicker iterations and reducing barriers associated with traditional design
tools (Hsu, Yang, and Buehler 2022). Moreover, ChatGPT's ability to understand human design
intentions helps to translate the input into Building Information Modeling (BIM) and parametric
Generative AI-Powered Parametric Modeling and BIM for Architectural Design 1
models, highlighting the potential of the architectural design process.
In this paper, ChatGPT extends its capabilities in creating 3D architectural models beyond the
2D images and text generation in traditional methods (Zhou et al. 2023). The paper presents two
distinct implementation methods: ChatGPT in BIM with Dynamo, and ChatGPT API within
Grasshopper in Rhino. This integration enabled the generation of complex, information-rich, and
parametrically controlled 3D architectural models in both frameworks. This allowed us to shift
towards more interactive and adaptive design methodologies beyond conventional 3D modeling
methods in architecture pedagogies. Subsequently, the two frameworks proposed in ChatGPT are
integrated with the Veras, AI-based ideation and visualization tool. This integration allowed us
to work more efficiently and effectively in their rapid design exploration by receiving real-time
feedback and generating design alternatives. Furthermore, we could engage in a more iterative
and dynamic design process by evaluating multiple design options and selecting the best solution
to evolve our design.
This paper comprises two main sections. The first section presents a unique implementation
method that integrates ChatGPT and parametric BIM with Revit and Dynamo. It outlines their
advantages and unique contributions. The second section introduces another implementation that
combines ChatGPT API with parametric modeling tools Rhino and Grasshopper. Both
workflows leverage Veras, for AI-based ideation and visualization. The paper discusses the
potential benefits and compares the experimental results.

2. ChatGPT for BIM Dynamo + Veras

In this section, we explore the implementation of ChatGPT (version 4.0) and Veras in the BIM
Dynamo design process. The integration of ChatGPT with Dynamo allows for an intelligent and
interactive design process, while the addition of Veras provides valuable feedback to refine the
generated design solutions (See figure 1).

Figure 1. ChatGPT + BIM Dynamo implementation workflow

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 2

The process begins with discussing the design task. ChatGPT helps identify the required steps to
accomplish the task, leading to the development of Python code as functions in the visual
programming environment Dynamo for the BIM authoring tool Revit. The Python code in
Dynamo is tested to ensure the design process works as intended. The Python code or process is
refined if needed. Once the design results are satisfied with the users, the chat with ChatGPT is
terminated.

2.1. Step-by-Step Conversations

This section analyzes the interaction between the user and ChatGPT in the design process. The
chat history is essential to understand how the AI understands the architect's intentions, converts
them into specific instructions, and eventually creates a BIM model of an example building
design.
● Step 1. Initial Query and User Input
This phase marks the start of the design process, where the user inputs the initial design prompt.
The initial prompt could be either a vague architectural concept or a more specific design
requirement. We began the interaction with a specific requirement to create a wall with a window
in a Dynamo (Figure 2).

Figure 2. Initial chat in the experiment (ChatGPT exported data)

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 3

 ● Step 2. User Feedback and Troubleshooting
Verifying and refining the generated Python code is crucial. The authors encountered bugs when
running the generated Python code in Dynamo and needed to refine the provided solution.
During this phase, the user provides feedback on the code generated by the AI model, which
leads to troubleshooting and code refinement. The process is that the user copied the ChatGPT
generated Python code into Dynamo, ran the Dynamo program, and normally received error
messages. The user then manually typed the error messages back into ChatGPT, which then
debugged the code and provided new code (See Figure 3).

Figure 3. Feedback and Troubleshooting

The user then re-copied the new code and replaced the old one in Dynamo, ran the Dynamo
program again, and normally the previous bug was resolved, but a new bug appeared. The user
then repeated the above process, until all the bugs were resolved (See Figure 4).

Figure 4. Debugging Interaction Cycle

Finally the working Python code was completely generated by ChatGPT. Also the debugging is
completely done by ChatGPT, except that the user had to copy (type) the error message manually
from Dynamo/Python into ChatGPT. The authors expect that this manual work could be
automated in the future by integrating ChatGPT into Dynamo through ChatGPT Application
Programming Interface (API). Therefore ChatGPT could provide troubleshooting and code
modification guidance, making it a versatile tool for interactive problem-solving.
Generative AI-Powered Parametric Modeling and BIM for Architectural Design 4
 ● Step 3. Chat for Creating BIM
It started with creating a wall with a window, then progressed to creating a room with four such
walls connected, and finally to creating a roof and a door. Each step was built on top of the
previous step’s outcome. Each step included a Python node with the ChatGPT-generated and
debugged code.
This approach allows for building a complex model from simple components step by step. Figure
5 shows the chat for the step of creating the room, after creating a wall with a window.

Figure 5. The user asked ChatGPT to create a room, and ChatGPT responded. (Data
exported from ChatGPT)

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 5

 ● Step 4. ChatGPT Guiding the Design Process
In this step, ChatGPT takes on a more directive role, going beyond the mere execution of
commands. ChatGPT could suggest general modeling processes or even recommend
modifications based on its understanding of the chat history (See Figure 6).
We requested ChatGPT to add a new door to the wall provided. ChatGPT utilized code derived
from previous prompts rather than creating a completely new process, with modeling steps based
on all previous conversation information.

Figure 6. ChatGPT suggest modeling process with modifications

This phase implies an enhanced level of interaction between the user and the AI model, which
uses its understanding of design processes and problem-solving strategies to actively guide the
user through various stages.

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 6

 ● Step 5. Interactive Developments
ChatGPT's capabilities extend beyond fixing errors to include adding new elements or expanding
the model. In Figure 7, we requested to add a door and roof to the existing model. ChatGPT
provided the updated Python code with four interconnected walls, windows, a door, and a roof
additionally.

Figure 7. Building Expansion and Core Updates

This demonstrates ChatGPT's flexibility and adaptability to the evolving design needs and goals
of the user. Thus, ChatGPT plays a significant role in the iterative and dynamic process of
building design, adjusting its responses and code based on user input and feedback.

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 7

 ● Step 6. Implementation and Fine Tuning
After generating each code with ChatGPT, users must instantiate it in Dynamo by creating a
Python Script node and entering the AI-generated code. Successful implementation of the code
requires user engagement in connecting the corresponding input and output nodes in line with the
script's requirements. In this experiment of creating the room in Revit/Dynamo, the AI-generated
Python code requires input nodes such as wallType, windowType, length, width, wallHeight,
windowWidth, windowHeight, and level. Users must determine the specific data type of each
input (i.e. a number, building object type, or level) and set the corresponding conditions
accordingly (See Figure 8).

Figure 8. The Python Script node and other input nodes created by the user manually
(left) based on the ChatGPT-generated Python code (right), in Dynamo, which can run to
create a building model.

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 8

This demonstrates the potential of ChatGPT in facilitating the architectural design process and
shows how AI can enable users to effectively and efficiently explore and realize their design
ideas (See Figure 9).

Figure 9. Final version of Dynamo (left) and ChatGPT-generated BIM (right)

2.2. Integrating Veras for Design Ideation and Visualization

This section discusses the integration of Veras, an AI-powered visualization add-in for
SketchUp®, Revit® & Rhinoceros®, that uses the 2D views of modeled 3D geometry as a
substrate for creativity and inspiration, into the ChatGPT + BIM process to facilitate the design
process. The initial model was a simple box-shaped space featuring four walls, a window on
each wall, a roof, and a door. While this serves as a useful starting point, initiating a new design
task from this simple model, particularly with limited design experience or unfamiliar tools, can
be arduous and exhausting. A significant challenge lies in the explicit semantics of design
elements such as walls, windows, doors, roofs, etc., as required by Revit for BIM. This requires a
basic understanding of architectural principles, posing a significant barrier for students and
designers new to these platforms.
Veras facilitated the generation of diverse design visualizations via user prompts, simply by
specifying the material and purpose of the building. For instance, Rendering A in Table 7,
prompted with "Wooden retail store", yielded a unique wood pattern direction for the walls and
doors and suggested fluctuation in wall sections, providing a information for subsequent design
and modification whereas Rendering B, prompted with "Concrete Office Building", reflected
dark glazing and setback openings.
This feedback loop extends beyond exterior design to interior spaces. For instance, Rendering C
& D(for a different BIM model of the authors), prompted with "Interior contemporary living
room with white paint and wood as major material", proposed filling the sunken living space and

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 9

replacing the TV wall with an open design, and removing the skylight windows for a flat finish
(Table 1) or keeping the skylight, depending on the material applied to the ceiling surface.

Table 1. Based on the views from BIM (left column), and the user’s prompts, Veras
generated different designs’ visualization (middle and right columns)

3. ChatGPT API in Rhino Grasshopper + Veras

This section explores the integration of the ChatGPT API with Rhino Grasshopper to generate
3D architectural models using natural language input. We discuss integrating the ChatGPT API
with Grasshopper and generating prompts as components.

3.1. Import ChatGPT API in Rhino Grasshopper

We incorporated a trigger component and data recorder in Grasshopper to create an iterative
design process. The trigger component facilitated the automatic and continuous execution of the
ChatGPT API requests. By setting the trigger interval (e.g., 2 seconds), we ensured the API
request would prompt every interval, allowing us to generate various designs rapidly.
While the API requests were created in every interval, the Grasshopper script also constructed
new coordinates per iteration, which allowed us to generate new geometry with each trigger.

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 10

3.2. API as Components
Our objective was to create actual geometries without text, human manual copy and paste, or
additional manual cleanup processes. We used 'loft' at the end of the process to generate the final
geometry. Initially, we attempted to create polynomial curves using the prompt: "Create one
polynomial with three variables x, y, and z." We observed that the generated equations included
numbering and text, which was undesirable for our purposes since we needed only equations
(See Table 2). Therefore, we added two additional prompts: "Do not number the equations" and
"No text, only equations."
Table 2. Prompt developments process

To create more consistent, regular, and patterned curves, we introduced constraints to generate
trigonometric polynomials using only sine and cosine functions. This approach led to the
successful generation of polynomial curves, avoiding drastic irregular geometries. To further
examine ChatGPT's understanding of shape descriptions, we tested again based on two different
prompts: 1) "Generate a polynomial curve that has placid, calm, and linear waves." and 2)
"Generate a polynomial curve that has surge, drastic, and crazy fluctuation waves".
Figure 10 shows the results regarding ChatGPT API generating two distinct shape designs easily
recognizable by humans.

Figure 10. Descriptive language implementation in ChatGPT API

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 11

This outcome demonstrates that ChatGPT can interpret human language input and generate
designs that cater to specific requirements or emotions. It highlights the importance of prompt
for unlocking the full potential of ChatGPT-generated architectural design, ultimately expanding
the possibilities in design methods.

3.3. ChatGPT as Generative Design Tool

We further investigated ChatGPT's ability to adapt to regulations or constraints in the design
process. We aimed to create closed curves that could be lofted together to form linear buildings.
We prompted the API to generate 10-sets of coordinates for a curve in the xz plane, keeping the
y-values constant at 0 to maintain alignment. However, this approach led to some twisted curves,
which resulted in conflicts when lofting the shapes together (See column a in Table 3).

Table 3. Prompt with topological condition

Therefore, we revised the prompt to focus exclusively on generating curves, including phrases
such as 'ensuring convex' and 'avoiding intersecting lines'. This refinement allowed ChatGPT to
produce a range of dynamic, non-intersecting, and convex (or concave, depending on the
viewing direction) shapes. The API adapted to these new constraints quickly and effectively,
highlighting ChatGPT API's potential in creating diverse and complex geometric forms, adhering
to specific design constraints.
Then, we introduced additional parameters for column generation. We stipulated that the
generated curve must always incorporate a circle with a radius of 6 units, serving as a structural
minimum cross-section framing to ensure stability. Simultaneously, the center point of the
interpolated convex curves could vary, but it had to be contained within a circle of radius 3,
centered at the origin. This introduced a randomized but controlled element in the design
process, maintaining structural constraints. This process shows its versatility and potential in
creating designs that satisfy practical architectural requirements.

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 12

3.4. Veras Visualization
Integrating the Veras allowed for rapid exploration of architectural style, locality, scale,
materiality, and context.
Initially, a building was rendered in Veras without specifying any user prompt. This resulted in a
building with an aluminum skin set against a broad, flat terrain with a sunset background (see
rendering A in Table 4). Intriguingly, Veras autonomously introduced a column to the lower
portion of the lifted mass, providing structural supplements.
Table 4. Veras renderings for design exploration

Subsequently, we specified 'Black metallic building' as a user prompt to evoke a darker, heavier
architectural aesthetic (rendering B). Veras responded by creating a building with a more
abstract, thicker skin, with the edges of the building clearly defined. An entry was introduced at
the lower center of the building, the only section in contact with the ground, further contributing
to its integrity.
Generative AI-Powered Parametric Modeling and BIM for Architectural Design 13
Inspired by the works of renowned architects like Zaha Hadid and Frank Gehry, we prompted
Veras to create buildings in their styles (rendering C and D). The results presented the ability to
emulate distinct architectural styles, generating designs of famous architects' iconic works.
Veras's adaptability was tested by inserting locality-specific keywords into the prompt. For
instance, based on the keyword 'Building in downtown Los Angeles', Veras created a background
of LA for a concrete and metal-skinned building (rendering E). Similarly, a prompt of 'Zaha
Hadid building in Downtown NY' resulted in a Manhattan-like background with a cool-toned
image rendering (rendering F).
Scale features were also explored (rendering G). The keyword 'pavilion' shrunk the scale by
changing the entourages’ scale and focused on the building's external skin and structural details,
such as stair-like features in the openings and structural elements supporting the lifted mass.
Finally, regional and material-specific factors were assessed. The same 'Pavilion' material was
applied to different regional contexts, such as Africa and the Arctic (rendering H and I), and each
rendering responded distinctively to its context. When a pavilion was rendered with the prompt
of 'Africa', it suggested a bamboo structure with wide spacing for ventilation and airflow. On the
other hand, in the Arctic, the bamboo spacing was rendered densely to imply a more compact
structure, with snow accumulated on its surface.

3.5. Veras Design Feedback

The ability of Veras to provide immediate visual feedback proved to be an effective tool in the
iterative design process. It was particularly useful in refining and realizing more complex
geometries generated through the ChatGPT API (See Table 5).
Table 5. Comparative Visualization of Design Iterations: basic prompt(left),
addition1(middle), and addition2(right)

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 14

This was observed when rendering the facade created by the ChatGPT API using the prompt
"Generate a polynomial curve that has surge, drastic, and crazy fluctuation waves" (See middle
column in Table 5). Veras translated this wave-shaped design into a rendering as a shell-shaped
roof. Although this result was intriguing, we wanted to create a more angular and protruding
poly shape output, which prompted a design revision.
We modified the instruction to the ChatGPT API by adding 'interpolated curve degree is 0'.(See
right column in Table 5). The intention was to emphasize sharp angles and discontinuities rather
than smooth curves. The API's response to this modified prompt was a markedly different model
featuring an angular, poly-shaped structure. The image produced portrayed a more fragmented
and jagged architecture, distinct from the initial shell-shaped roof design. This process highlights
the power of AI-based Visualization tools for rapid visual feedback and iterative design. Veras
effectively facilitated visualization in the design process by generating immediate designs
created by the ChatGPT, enabling designers to iterate and refine their ideas rapidly.

4. Results and Discussions

In the conventional design process, architects and designers have to imagine and model the
elements of the building model, such as walls, ceilings, roofs, and site models. They then need to
create renderings to visualize the design, get feedback, and refine the model, repeating this
iterative process until a satisfactory design solution is fulfilled. This repetitive process is not only
labor-intensive but also time-consuming. This can lead to increased project complexity and
budget, requiring constant collaboration and real-time updates from all team members.
Furthermore, this iterative modification significantly escalates the potential errors, posing
additional management, computational challenges, and complexity.
The novel design process proposed in this study integrates ChatGPT, its API, BIM, parametric
modeling, and Veras, resulting in a more efficient and creative workflow. This integration
reduced the labor and time typically consumed in the iterative design process for coding in
Dynamo. Furthermore, the ChatGPT API in Grasshopper generated codes and polynomial
equations based on descriptive language prompts, facilitating the creation of complex and
diverse geometric shapes. Subsequently, implementing Veras offered immediate and realistic
feedback on designs, enabling a quicker and more effective iteration process. The AI-based
rendering tool's ability to manipulate architectural style, scale, materiality, locality, and context
greatly enhanced the design exploration process. This significant shift in the design approach
reduced the time and complexity of creating and adjusting Grasshopper scripts, enabling
designers to focus more on creative exploration and less on technical scripting.
Our exploration of ChatGPT and Veras in architectural design applications reveals promising
potential but also uncovers critical challenges for further research. Specifically, it is still unclear
how well the current generative AI can address complex, bespoke, or customized space design
Since these design intricacies are crucial to generative design (Nagy et al. 2017), a more
comprehensive study into how AI can effectively accommodate them is imperative. Furthermore,
we cannot overlook the physical feasibility and structural integrity of AI-generated designs. The
Veras implementation process highlighted some cases where AI-generated designs resembled

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 15

illusionary constructs like Penrose's Stairs (Nakatsu, Takahashi, and Moriya 2013). Although
visually impressive, these designs were architecturally infeasible, reminding us that the boundary
between perception and reality in AI-guided design requires further exploration.

5. Conclusions

The journey towards realizing the full potential of AI in architectural design is ongoing. This
study has demonstrated that integrating technologies like ChatGPT and Veras into the
architectural design process can enhance efficiency, collaboration, and creativity. However, this
new frontier in architectural design presents challenges that require continual exploration and
refinement. As AI technologies continue to evolve, their capabilities for generating unique and
novel design solutions will also increase, pushing the boundaries of what is conventionally
considered feasible or imaginable in architectural design. Despite the immense potential, it is
crucial to recognize the challenges associated with these technologies. The limitations identified
in this study should serve as catalysts for future research and development efforts. As we
navigate the evolving landscape of AI in architectural design, a continued scholarly and
industry-wide dialogue is critical. This dialogue, guided by rigorous empirical research and
inclusive professional discourse, will be instrumental in fully harnessing the potential of AI in
architectural design.

6. Future work

There are many potential research directions based on our study. For example, while Veras can
generate many new design ideas and their visualizations as 2D images based on the Revit or
Rhino models’ 2D views, future research project will investigate the reversed process: how to
create 3D parametric BIM based on the new design ideas generated by Veras or other AI tools?
As a final experiment for this paper, we thought: why not ask AI to help us look into the future?
When we asked ChatGPT on May 24, 2023: “What is the future of generative AI for
architecture?”, here’s what ChatGPT responded (Figure 11):

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 16

 Figure 11. GhatGPT Response about the Future of Generative AI for Architecture
Generative AI-Powered Parametric Modeling and BIM for Architectural Design 17
We think that ChatGPT responded well to the major future research directions. Currently
researchers in related fields could tackle the challenges in these directions. The authors are
working in some of the directions, including optimized design, human-AI collaboration, Virtual
Reality and Augmented Reality, etc.

References
Cross, Nigel. 1982. “Designerly Ways of Knowing.” Design Studies 3 (4). Elsevier: 221–27.
Eastman, Charles M. 2011. BIM Handbook: A Guide to Building Information Modeling for Owners,
Managers, Designers, Engineers and Contractors. John Wiley & Sons.
Hsu, Yu-Chuan, Zhenze Yang, and Markus J. Buehler. 2022. “Generative Design, Manufacturing, and
Molecular Modeling of 3D Architected Materials Based on Natural Language Input.” APL
Materials 10 (4): 041107. doi:10.1063/5.0082338.
Nagy, Danil, Damon Lau, John Locke, Jim Stoddart, Lorenzo Villaggi, Ray Wang, Dale Zhao, and David
Benjamin. 2017. “Project Discover: An Application of Generative Design for Architectural Space
Planning.” In Proceedings of the Symposium on Simulation for Architecture and Urban Design,
1–8.
Nakatsu, Kana, Tokiichiro Takahashi, and Tomoaki Moriya. 2013. “Animating Impossible Figure from
Twisted Torus to Penrose Stairs.” In SIGGRAPH Asia 2013 Posters, 1–1.
Negroponte, Nicholas. 1975. Soft Architecture Machines. MIT press Cambridge, MA.
Radford, Alec, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish
Sastry, et al. 2021. “Learning Transferable Visual Models From Natural Language Supervision.”
arXiv. doi:10.48550/arXiv.2103.00020.
Radford, Alec, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, and others. 2019.
“Language Models Are Unsupervised Multitask Learners.” OpenAI Blog 1 (8): 9.
Ray, Partha Pratim. 2023. “ChatGPT: A Comprehensive Review on Background, Applications, Key
Challenges, Bias, Ethics, Limitations and Future Scope.” Internet of Things and Cyber-Physical
Systems 3 (January): 121–54. doi:10.1016/j.iotcps.2023.04.003.
“VERAS.” 2023. EvolveLAB. Accessed April 30. https://www.evolvelab.io/veras.
Zhou, Ce, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, et al. 2023. “A
Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT.”
arXiv. doi:10.48550/arXiv.2302.09419.

Generative AI-Powered Parametric Modeling and BIM for Architectural Design 18