AI in engineering is not new. Although the name has changed, we’ve used narrow AI for things like object recognition in automation lines or generative design for a long time. What is new is the influx of tools claiming to fundamentally change the way we work. I’ll try to break down what I’ve learned as mechanical engineer into three sections: AI CAD tools, LLMs like OpenAI’s ChatGPT, and text-to-3D model generation.

AI CAD Tools

Depending on your (or your company’s) preference, both Autodesk and Solidworks have developed tools to help automate repetitive tasks for both part and assembly creation. Although they may seem a little underwhelming at first, the ability to quickly create hole patterns or fully defined mates can save you hours or even days.

What excites me more is the promise of offloading some of the tedium of creating drawings. Autodesk has introduced Drawing Automation, which can generate the layout of a part using your templates and even create dimensioning strategies. Since my company uses SolidWorks, I will have to wait until they follow suit.

These are undoubtedly the first steps in a series of updates we will see from CAD companies as they begin incorporating AI into their software suites.

Large Language Models (LLMs)

Using LLMs or other AI tools may not be as straightforward for a mechanical engineer as it is for a software engineer, but I believe that even mechanical engineers will need to adapt by learning to use AI tools effectively. This means focusing our efforts on areas where human expertise, intuition, and judgment are critical.

I posed some test questions to OpenAI’s ChatGPT (ChatGPT-4o) to see how it measured up. One design constraint I have been grappling with lately is how to isolate pogo pins on our PCB from the aluminum housing in an extremely sensitive circuit. I had previously researched machinable materials with low dielectric constants and found that Teflon or PEEK would be our best options. Not only did ChatGPT identify the same materials (results below), but it also suggested some ceramics that would be good candidates for high-temperature applications.

Although ChatGPT is the dominant player in the field, there are startups developing LLMs specifically for mechanical engineers. Similar to GitHub’s Copilot for software engineers, Leo was “trained on a dataset of millions of man-made products and uses machine parts as tokens (bolts, bearings, etc.), combining them into DFMA-compliant products.” The hope is that a model trained with parts will provide better insights into design constraints. I haven’t tried Leo myself (as there is no free version), but perhaps in the future, I will purchase a month to trial it.

Ultimately, while there is great promise in using LLMs for design work and research, an experienced engineer is still needed to double-check the AI’s work. You still need to be able to justify or understand whether the AI’s suggestions make sense.

Text-to-3D Generation

The last topic I’ll touch on is probably the farthest in the future. Text-to-image models have been around for quite some time, but what I’ve been eager to see is text-to-3D modeling. Autodesk has a highly experimental text-to-3D model called Bernini, which may someday be able to produce parts from a text description, but it’s not yet ready for prime time.

The closest solution I’ve seen is Meshy. Meshy can take either text or a 2D image and attempt to create a 3D model—sounds perfect, right? Almost… These models are polygonal, which limits their usefulness for production parts. Still, they can be generated with a texture map and may be useful for MVP concept generation, not to mention video games and animated movies.

Prompt: a sci-fi scientific instrument with white sides and black trim with a touch screen on the front, a door beneith the touch screen, and three bottles on the side

Conclusion

Although not as mature as AI tools in the software design space, tools specific to mechanical engineers are being developed. Generalized AI tools like ChatGPT can still be a great resource but require an engineer with a highly tuned “B.S.” detector. Unfortunately, we will still need to make our parts the old-fashioned way for the foreseeable future.