CAD and AI
AI is actually in use in a lot of places. In writing it is regrettably used frequently. I’ve tried it out, and it has an easily identifiable tone. The content is pretty generic, and it avoids giving you the kind of information you generally read articles to get. AI is also in use for making images, video and audio. This is not the future or speculation, this is now. Of course the images can be as surreal as the writing. To the right is an image I got from AI when I was asking for a photo-realistic image of a trombone in the grass. If you know much about trombones, you know the result shown is more of a twisted nightmare than realism, photo or otherwise.
Not all AI is this ridiculous. The writing is actually pretty good if you’re just looking for generalizations, or to reorganize something you’ve already written. We’ve all heard AI generated music that is imaginative in a way we’re not familiar with, or something that imitates an artist that’s no longer with us like Johnny Cash. It’s maybe unfamiliar and possibly strange, but believable.
One of the things you might learn about AI from using it is that the old computer wisdom of GIGO (Garbage In Garbage Out) is still in effect, even with AI. Part of the secret to actually using AI (as opposed to building or training AI) is that you have to give it some sort of prompt. You want it to write a term paper on the Civil War? You have to be able to give it information about what you think the result should look like. You want to create a great picture of a trombone in the grass? You have to feed it good information, and prompt it properly to do what you want it to do.
But that’s not why you’re here. You’re here to read about how AI might impact mechanical design and CAD. For the time being, AI is mostly having an impact in marketing.
I wrote about this topic 5 years ago (https://dezignstuff.com/the-role-of-ai-in-design/), where I included a portion of an interview with John McEleney where I asked a question about AI. All the answers were sufficiently vague at that point, as it was still very early on for AI. At least now we have a glimpse at the kinds of things it can do.
The big aspect of AI that design tools will have to deal with is that you need a lot of data to train your AI. Well, we’ve got that. Data, yes. Real information, maybe not. We would have to put all of that data into some sort of context for AI to get much benefit from it. For example, sort info into different types of design. Machine design might be the most obvious place to get started. AI should be able to be trained to design specific types of machines or at least different types of mechanical elements.
For example, to get started we should be able to train AI to design say a hinge for different types of applications. This kind of AI help can’t be that far away. Functional design. A hinge is a simple mechanical element, and if AI can understand how to write English sentences, it should be able to make a hinge to support a door in a house, or a car door, or a bank safe vault door. Maybe a connection between components, joints, wheeled mechanisms. Mechanisms like this are things that engineers redesign a hundred times a day, and we might be able to save some time and avoid repetitive tasks to hand some jobs like this over to the computer.
From there things like mold design where we have engineers do the same sort of design over and over again should be able to be automated. Machines designing machines.
Maybe it graduates from there to understanding the design of an electrical motor – to make a smoother, more powerful, more efficient electrical motor that runs on lower voltage. At this point it would be improving on things that humans have been doing. Maybe the computer winds up teaching us how to make certain types of decisions that we haven’t been able or willing to make in the past.
But all of that is just the design of small components of systems. I think in the short term, less than 10 years, we can have an AI system that essentially does the work that software like DriveWorks or Rule Stream does for us now. These are configurators or engineer to order systems, where you set up a parametric model, and then write a program to create variations of that for you. Like to simplify it, say you want to order a sandwich, and you order a 12″ roll, turkey, tomato, lettuce, mustard, Gouda cheese. And first the AI designs it for you, then the robotic system builds it.
Rule Stream demos used to center on something like a chemical processing plant with tanks and piping, all customizable within certain parameters. AI should be able to take over the programming side of this for us. We train it on what we want, what changes need to be engineered, feed in the parameters, and it kicks out a set of drawings.
The future where Star Trek TNG used voice commands into the computer to reverse engineer apparatus from a nightmare is a long way away, but maybe not as far as we think. Once we get the pattern of how to connect AI to design, and AI understands the language of engineering, we are most of the way there.
Generative design might a partial answer of how “AI” could be useful for mechanical design
Yeah, generative design. Another borderline blue sky tool. I think there has to be a way to take generative design less literally. As it is, it tends to grow roughly organic shapes, with “roughly organic” not being a compliment. If it could find a way to make the shapes look more intentional – more actually designed than just allowed to happen, that would be great, and that might be something AI could do. But is it AI or just better programming? Still, you have to give it input. I kind of consider generative design to be a form of AI already. Your input is the set of boundary conditions for a stress analysis, and it outputs a geometry to minimize the stress.
I recently learned of a company working at outputting prismatic shapes from generative design, suitable for CNC machining. InfinitForm is currently in closed beta at https://infinitform.com/.
Generative Design is essentially FEA in reverse, so straight algorithms to figure that out. The place for AI would be after that to do what you are referring to which is make it an easily manufacturable part using the manufacturing method and material of choice (machined, molded, folded sheet metal, etc…).
I would be happy if AI could just perform the relatively simple and tedious tasks like reverse engineering an editable BREP model from a point cloud/mesh model or analyzing model tolerances against manufacturing method for costing/manufacturability. I’m with Ralph on this… A lot of what is called AI is really just the “fuzzy logic” of the 90’s.
I think the original AI was CAM software. You put in a drawing or model and it gives you back toolpaths, a painful and laborious process to write by hand. I read a history of 3D CAD where aerospace companies were willing to pay 100s of K for one of these early workstations.
I think there is a huge gap in what so-called AI can do now, and what we wish it could do to help take over tedious tasks:
Step 1: AI is amazing!
Step 2: ???
Step 3: AI is actually useful.
As you note, AI is mostly used by marketing, and AI firms are desperate to maintain their valuations by promising AGI RSN (real AI, real soon now).
I think that what CAD vendors call AI is what I would instead call “search and replace” or “most-recently used” or “rules-based” or “macros.” Those tasks require no thinking by the software, because they depend on pre-determined patterns; the thinking has been done by programmers writing the code.
I am ready to call it AI when software can handle problems that don’t fit predetermined patterns, such as when I edit a badly-written paper that contains illogical sequences of arguments, spelling so poor the spell checker cannot figure it out, and incomplete/run-on sentences.
Thank Matt, it is incredible to read this.
Talking about AI in CAD makes me remember scene from Iron Man, when Tony Stark asked Jarvis to design his power suit.