
Generative Manufacturing - Computer Aided Design - Approaches and Challenges.
Explore the world of Generative CAD: using Generative AI to power Computer Aided Design workflows. This blog will walk you through current approaches, from code generation using LLMs to specialized model, and MCP based approaches, and highlights outstanding challenges and considerations.
Introduction
The Fourth Industrial Revolution promised major changes in how industry and society operate, from artificial intelligence, to 3D printing, to gene editing, to advanced robotics, and the ubiquity of IoT devices, we can see the results of this revolution all around us.However, hidden away in the background are the engineers, the architects, the designers, toiling away, creating the data centers that power the AI revolution, designing the printers, the robots, ans all the other hardware of modern society, that we all take for granted.
From the world: we thank you for your service.
For a long time these magnificent humans would toil away with paper and pencil creating designs which could then go on to be manufactured: blueprints.
As time moved on, digitization moved in, and helped removed the backbreaking work of designing. CAD (Computer Aided Design) software is just how engineering and architecture are done now, the improvements in speed and accuracy that CAD offers is just too good to ignore.

However, the time spend with CAD is still labor intensive, Generative AI offers an opportunity to reduce the time spent by engineers creating models, freeing up time to work on the harder problems in the engineering workflow, while democratizing access to 3D workflows for those of a less technical background.
But, how does this all work?
In this blog we'll walk through the current approaches to Generative Design, the pros and cons, and how to get started with each. By the end of this blog, you should have a good understanding of how to implement the approaches, and hopefully an appreciation of how you can fit these into your workflows.
What is Generative Manufacturing?
Generative manufacturing applies the power of artificial intelligence to generate and execute optimal solutions given customer-defined constraints and parameters, such as functional specifications, cost, and lead time.Generative Manufacturing is general term and can be applied to the whole product life cycle, from design and prototyping, to testing, and the manufacturing process.
This blog focuses on the Generative Design component of Generative manufacturing.
Generative Design
Generative Design , is, as the name suggests, the use of algorithms and artificial intelligence in to design of a component, product, build, sculpture, or an other physical object.
Generative Design is not a new concept. Antoni Gaudi's Sagrada Familia is an early, and probably the most famous example of using a rule based geometrical forms to create structures. More widely in the architecture space, Generative Design allows the architect to explore a wider solution space, providing the architect with more possibilities, and enabling more creativity.
On the engineering side Generative Design has been used to great effect in the design of sustainable solutions. This approach to Generative Design uses physical simulations and traditional algorithmic approaches and creates a feedback loop between the simulation and the design to test a wide range of options for their sustainability.
The GenAI approach to Generative Design is still in its infancy, and as such, is not yet at the level of optimizing designs, in the way the approaches discussed above can. Instead, genAI approaches focus on speeding up the design process. In future with the use of Agentic patterns , it may become possible to incorporate part of traditional Generative Design with GenAI to create powerful, fully automated solutions, but not yet.
Why does any of this matter?
This is a great question, why does this matter? Well, design is hard, and creating the CAD models is time intensive, anything that can speed up the process can result in getting new products to market earlier, and allows the designer to focus more on the more complex challenges in the design workflow.Just as CAD freed up time and allowed the design of bigger and more complex solutions, by bigger teams of designers, Generative AI provides the possibility to fundamentally change the design process in the same way: opening the door to bigger, more complex solutions, in less time, and at lower cost.
Current Approaches
There are currently 4 main approaches to use Generative AI to create physical objects. For each of these approaches we'll dive into what they are, the challenges, and a quick demo. For these demos we're sticking to open source tools. That means some of the tools we use aren't strictly engineering CAD packages: Looking at you Blender. But they allow us to demonstrate the theory, and the ideas and techniques are directly transferable to more engineering focused tools.
Text-to-Parametric
Parametric CAD software, for our purposes lets consider openSCAD defines a model as a script, or history of simple geometric shapes, and changes to those shapes. Think of dimensions of a rectangle, or the angle between 2 lines. In the case of openSCAD the software is entirely script driven. This allows standard LLMs, such as Anthropic's Claude or the Amazon Nova family of models to directly create a CAD design which can then be used by the software.
This approach is by far the easiest to begin with, all you need is access to an LLM and you can begin. There is even an opensource solution available ( CADAM ) to help you start, if you like the openSCAD approach. Longer term, this approach would likely benefit from model fine-tuning.
The biggest challenge with this approach is specifying what you need to natural language: as this uses lengths, angles, etc. to define a geometry, if you do not specify what you want, the model will make that decision for you! This can lead to very long prompts, sometimes longer than what you would need to write using openSCAD directly, although pure natural language does lower the barrier to entry for systems such as these.
Getting Started
- Head over to OpenSCAD and follow the instruction for your operating system.
- Open up your preferred LLM chat interface
- For the demo below, the Amazon Bedrock Chat / Text playground was all that was needed.
- If you're new to Amazon Bedrock, take a look at the Getting started with Amazon Bedrock guide.
Demo
For demo purposes I chose to Claude 3.7 Sonnet to create a simple bracket.
Bracket can mean many things, for this demo, lets try to recreate something like this:

Prompt
create an openSCAD script to create a bracket. The bracket should have a base 15mm x 15mm and 5mm thick. Centered on this base should be a protrusion that is 10mm tall, 5mm wide, and 10mm long. There should be a hole with a radius of 2.5mm through the long face of the protrusion offset 5mm from the top of the protrusion. The top edges of the protrusion should be rounded.Output
// Parametersbase_width = 15;base_length = 15;base_height = 5;
protrusion_width = 5;protrusion_length = 10;protrusion_height = 10;
hole_radius = 2.5;hole_offset_from_top = 5;
corner_radius = 1; // Radius for rounding th top edges
module rounded_cube(size, radius) { hull() { // Place spheres at the top 4 corners to create rounded edges translate([radius, radius, size[2] - radius]) sphere(r = radius); translate([size[0] - radius, radius, size[2] - radius]) sphere(r = radius); translate([radius, size[1] - radius, size[2] - radius]) sphere(r = radius); translate([size[0] - radius, size[1] - radius, size[2] - radius]) sphere(r = radius); // Add a cube that's slightly smaller in z to create the bottom part translate([0,0,0]) cube([size[0], size[1], size[2] - radius]); }}
// Create the bracketdifference() { // Base union() { cube([base_width, base_length, base_height]); // Centered protrusion with rounded top edges translate([(base_width - protrusion_width) / 2, (base_length - protrusion_length) / 2, base_height]) { rounded_cube([protrusion_width, protrusion_length, protrusion_height], corner_radius); } } // Hole through the protrusion translate([base_width / 2, base_length / 2, base_height + protrusion_height - hole_offset_from_top]) { rotate([0, 90, 0]) cylinder(r=hole_radius, h=protrusion_width + 2, center=true, \$fn=32); }}Using this output is as simple as pasting the output into OpenSCAD and clicking render!
If you are following along at home, that failed, with this error
ERROR: Parser error: syntax error in file, line 52 . Claude adds an unexpected \ before the function call on that line, remove it and it will render correctly, and you should see this:
This is a solid, ready to manufacture thing. You can export this object to an stl file and print it with no issues. In fact, just for completeness you can see the print below:

Model Context Protocol (MCP)
MCP approaches use the API of existing CAD software and allow standard Large Language Models (LLMs) to interact directly with the CAD tool. For our example, we'll step away from "real" engineering tools and use Blender , an open source modelling tool, that is designed for artists, but has found great use in the 3D print community.
This approach is my preferred solution when tools expose an API that allows manipulation of objects, although it may be the most finicky approach right now, the control it provides and instance feedback make it worthwhile.
MCP is the standard for interacting with tools via an LLM. This is great, it means as a builder, there is less worry about custom syntax and model fine-tuning, it, mostly, just works out of the box.
This approach is also the most seamless for human in the loop design, as its interacting with your favourite CAD tool, you can watch it work, and jump in and fix or alter the Generative output in your tool, no need to load an output file, or import a mesh.
The biggest challenge right now for this approach is some CAD tools lack a usable API, and the need to create an MCP server to front that API. If the MCP doesn't exist you'd need to create it!
Getting Started
If you're new to Amazon Bedrock, and skipped the Text-to-Parametric approach above, take a look at the Getting started with Amazon Bedrock guide.
Install Blender
Setup your coding environment
Hopefully you already have an IDE with an Coding assistant setup. If not. For this demo Visual Studio was used with the continue.dev plugin, using Claude 3.5 Sonnet from Amazon Bedrock.
- To install VSCode go to VisualStudio and follow the instructions for your operating system.
- To install the continue.dev extension open VSCode, click on
Extensionsand search forContinue.continueand install theContinue - open-source AI code agent
- To configure continue.dev to use Amazon Bedrock, update the main
config.yaml(found at~/.continue/config.yamlif your on Mac or Linux) file to this:
name: Blender Agentversion: 1.0.0schema: v1models: - name: Claude 3.5 Sonnet provider: bedrock model: anthropic.cluade-3-5-sonnet-20241022-v2:0 env: region <your bedrock region> profile <your AWS credentials profile from ~/.aws/credentials> roles: - chat - edit - applyAdd the MCP server to Continue.dev
- Ensure the
uvpackage manager is install. If not, follow the instructions for installing uv
- In the Continue.dev menu in VSCode click on Tools, then click
Add MCP Server. This will open a new.yamlin VScode for you. Update that file to look like this:
name: blenderversion: 0.0.1schema: v1mcpServers: - name: blender command: uvx args: - blender-mcp env: {}Install the Blender-mcp add-on to Blender
- Download and unpack the blender-mcp package from blender-mcp
- Open Blender
- Go to
Edit > Preferences > Add-ons
- Click the dropdown arrow, then click
Install from disk
- Select the
addon.pyfile you just downloaded
- In blender go to the 3D View sidebar (press N if not visible)
- Click the
BlenderMCPtab
- Click
start mcp server
Demo
As before we're going to attempt to create a little bracket.
Spoiler Alter: It did not go as planned
Prompt
using blender create a cube that is 15x15x5, extrude this cube to make a protrusionthat is 10x5 and 10 tall. Finally add a hole through the long (10) side of the protrusion that has a radius of 2.5 and is 5 from the top of the protrusion.Output

The output is created directly in the Blender interface, so no copying and pasting here. Its just ready to be manually updated, or exported and printed. With the MCP approach you will not get the same output seen below, you'll get something different, any potentially magical.

The object is well designed, no issues with edges, no overly complex geometry.
print came out well too:

Well clearly this is not the bracket hoped for. This is, well I don't know what it is. Maybe it's a model of the headquarters of some supervillain, or maybe its part of the conning towers of a space ship from a grand space opera, or, maybe I just love the Venator Class Cruiser too much and see it everywhere.

There was a desire to continue refining the output until it matched the text-to-parametric output. But the desire to keep to a single prompt and the interesting issues with the Blender MCP this output highlights won out.
FYI: For those interested in what this could have been, I've added a human in the loop output at the end of this blog.
What are these issues?
Well, there are many ways to create the bracket we wanted in Blender, you could:
- create a cube (15x15x5), loop cut it and extrude the center
- create a cube (5x5x20) and add smaller (5x5x5) cubes to either side
The Blender MCP struggled with these approaches and settled on simply extruding the full top face and shrinking the top to be the correct dimensions, leaving the pleasing pyramid shape.
Text-to-Mesh
Text-to-Mesh approaches use a Generative AI model, such as the Tencent HunYuan3D-2 model , that can take text or images as input and outputs a mesh.These models are typically marketed as 3D asset creation models, such as video game characters, and were not specifically created for engineering use case.
These models are diffusion models, and are easy to get started with. You can think of them as 3D versions of stable Diffusion . That simplicity comes with a cost. The models are designed to create 3D meshes that look good, just like Stable Diffusion is designed to create images that look good. They don't have any real understanding of what a mesh is, or what is required to create a physical object. What does this mean? Well, the objects will not be dimensionally correct, may have holes, or be made of multiple disconnected parts, and most of all, these models create complex, often non-manifold, outputs.
Getting Started
For this demo a Amazon EC2 was used to host the model. If you are new to EC2 take a look at the Get started with Amazon EC2 guide.
This demo used a g6e.2xlarge instance.
To install the model follow the instructions in the Github repository . Using the Gardio App is highly recommended.
Demo
Prompt
male swivel base flange square base single protrusion with a single hole in itOutput
The output of these models is a mesh, so no text output here. Lets jump straight to the render. For this, we'll import the stl into Blender and take a look:

This looks okay, the threads on the mounting holes are clearly not real threads, and its not the part that we wanted, but we got something. As with the MCP example we could keep iterating, but this is good enough for our current needs, which is to understand the challenges of this approach.
To do that, lets take a look at the objects mesh:

Along with complex meshes, these models also tend to generated non-manifold objects (more on that in the Challenges section), but for now, just now that's an issue.
This mesh has the following unwanted characteristics:
- 3 non-manifold edges
- 12 Intersecting faces
- 6412 Zero faces
- 1 Zero edge
This is not an output you're going to want to manually alter later. The mesh is just too complex. However, we can try to fix this. For the demo Blender was used for this process, but this is not the only way to do this:

But when we look at the mesh, its going to be much easier to work with:

Your options for post-processing here are still limited to direct modelling, but is going to be easier after simplifying the mesh.
For those that like to see the finished product. Sorry, not printing this one
Mesh-to-Mesh
Mesh-to-mesh models are cutting edge, these models can take in text, images, or meshes, and output a mesh. SGS-1 is an example of such a model.
These models are designed for engineering and use a parametric approach to creating the output, which means it creates a mesh that should be more dimensionally correct than the standard artistic mesh models. The other major benefit is that these models can take a mesh as input, making it easy to update an existing mesh asset.The downsides of this approach is that is it still a research area, and the outputs are still meshes, which means you're really limited to direct modelling approaches if you want to manually alter the output.
Getting Started
Getting started with this type of model is still difficult, the SGS-1 model is the first generative model for structured CAD, and its not yet open, but the producers Spectral Labs does provide a form if you are interested in collaborating, or deploying the SGS-1 model. They also provide a playground to try out the approach.
Demo
Prompt
The playground is image to mesh only. So for this one, we'll simply upload the reference image.
Output
The output of this models is a Boundary mesh in STEP format. Lets take a look at the output:

Be warned, the output isn't a single mesh, its many meshes, like 1000's of meshes, this may have an impact when importing in to your favorite CAD tool.
For now, we can settle on exporting this as an STL, so we can move back to Blender and take a look at the mesh view:

As with the text-to-text ouptut this mesh has the following unwanted characteristics:
- 2774 non-manifold edges
- 58 non-contiguous edges
- 328 Intersecting faces
- 5940 Zero faces
- 4 Zero edges
Attempts to fix this mesh did not go so well. However, the output does show some interesting issues with image-to-mesh. The odd shapes at the corners appear to be caused by the bright reflection in those areas, whereas the missing section on the upright portion of the model seem to be cause by the shadow in this area.
Input quality clearly has an impact on output.
You might be thinking, that's not a great output, but if you've spend anytime with 3d scanning technology, this output is a game changer compared to the output meshes those tools create.
Current Challenges
These approaches are right at the edge of the map, and here be dragons! Although these dragons are relatively small, but its good to know what they are.
Non-manifold geometry
Non-manifold geometry is, essentially, a bad mesh. It might look nice, but its got some issues that make manufacturing, such as 3D printing difficult. Think of a manifold mesh as a mesh that encloses space, where this is a clear separation between the inside and the outside.
There are a host of problems what are technically not a non-manifold geometry, but which result in a similar set of issues, so I'll lump them all in here:
- An object where an edge connect more than 2 faces
- Disconnected vertices or edges
- Internal faces
- Areas with no thickness
This problem shows up with the approaches that create meshes: text-to-mesh, MCP (at least in the case of Blender), mesh-to-mesh.
The solution to this problem is to manually edit the output mesh to fix the issues, or to post-process the output using a tool to automatically fix the mesh.
As we are already using Blender for this simple example, this can be achieve in Blender using the 3D print toolbox, it's only 1 click away. If you use Windows on your workstation, Microsoft provides a mesh fixer, that works well.
Overly Complex Meshes
Again this effects the mesh approaches. This one is really simple, the outputs of the mesh models are overly complex, way too many vertices and edges that add nothing but complexity and difficultly if you want to alter the generated model.
There are solutions, again using Blender as an example, there is a simplify mesh function, that does exactly what it says on the tin.
Suitability for Post-processing
The suitability for post-processing is dependent on the approach taken.
Text-to-mesh and mesh-to-mesh are difficult to post process, beyond simplifying the mesh and making it manifold. These outputs are only suitable for direct modelling (think of the outputs as clay, you can push and pull it into shape, but not much else).
The text-to-parametric script approach is simple to alter, simply load the file and starting working as you normally would.
The MCP approach is the simplest to work with, as it is suitable for human in the loop, real-time collaborative editing with your AI colleague, the output is in the native format you want, and can be used just like any other model.
Defining an end-to-end workflow
Over the course of this blog, you may have came to the conclusion that these Generative Design approaches cannot be used in isolation, you'd be right. These are not quite JARVIS, maybe one day we can ask a computer for a titanium-Gold alloy super suit, but today is not that day.
Today these approaches need to fit into existing workflows, they contain the possibility to improve productivity, remove boilerplate and let engineers solve the real problems. But to do that these tools need to be as invisible as possible, using them should not incur extra overhead.
This is the big challenge, and like most big challenges in software, it's also the challenge with the biggest human component.
There is not solution here: this is a people and process problem. But there are suggestions!
Start Small
If you're looking to accelerate your, or your engineers workflow, get familiar with the MCP for your CAD tool of choice, and begin using it.
Start constrained, then add freedom
If automated workflows are your direction of travel. Using Generative AI with tight constraints works better than given the AI free reign.Consider that it is easier to have AI update the parameters of an existing model than to create a model.
Conclusion
Generative Manufacturing and Generative AI Design are new, and it will be some time before the space matures, however the possible value of these approaches is too high to be ignored.
Each of the approaches shown above come with their own challenges, and situations in which they excel, and this space is moving quickly, what is written above is likely to be out of date almost before it is published. This is a good thing.
What is shown above should give you enough knowledge to get started and to test out these approaches.If you create CAD software, start experimenting with implementing your MCP today.
If you work with CAD software, try out the approaches above, and start to think about how your workflow changes given an AI CAD buddy.
Human in the loop MCP output
Prompt
using blender create a male swivel base flange square base (15x15x5) with a single protrusion (10x5x15) with a hole in itOutput

Manually pulling the top cube onto the base and doing a quick remesh to join them together gave us a good output:

Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article