AI a golden opportunity for design software vendors. Hopefully, they don’t blow it.
In the late 1950s when computers were still new and CAD was merely a research project, CAD stood for several things: computerized design,
computer automated design and computer-aided design, depending on who was doing the research. The hyphen in computer-aided design is significant, argued MIT professor Robert W. Mann at a series of seminars in 1959.
It emphasised the computer is aiding the engineer. With no hyphen, computer aided design implied computers were doing the designing. Mann wanted it to be clear that computers, now able to draw lines and arcs, were going to help engineers and not replace them.
What was old is new again. Engineers again fear for their jobs — this time because of AI. It doesn’t help that the hyphen meant to reassure them has disappeared. With no reminder of CAD’s original intent as an assistant rather than a replacement, engineers might think their obsolescence was the plan all along.
But even then, engineers must admit if CAD, if it ever aspired to design, has failed to do so. It’s only ever helped humans detail designs; engineers have always had to think of that design first. Only now with AI can we dream of computers helping engineers design.
But let’s not get carried away and assume AI will design whole cars, ships, buildings, and bridges. It can’t and won’t ever. But soon it should be able to help engineers design these things. It could help just little at first. Here’s how:
1. AI should RECOGNIZE design intent.
Let’s say an architect is designing a home addition and making a floor plan. When drawing lines for walls, the CAD program should recognize the intention and convert such lines to walls complete with drywall, sheathing and studs 16 inches. The addition of a pitched roof should trigger details to automatically match the pitch to existing rooflines and joint the roof sections. It should calculate the joists and rafters necessary at the spacing required to satisfy local codes. AI should also be able recognize shapes from data input. It should be able to connect the dots, as it were, from point clouds and recognizes them as physical parts. Or match parts in a photograph with models in a library.
2. AI should SUGGEST shapes and parts.
No engineer should have to search for standard parts or shapes, and worse, upon failing to find them, give up and draw shapes or design parts from scratch. Engineers can get bogged down making simple parts, dimming any creative spark they might have had. AI should recognize their intent and suggest appropriate shapes or parts. If an engineer is laying wiring, AI should suggest a wire harness. If an engineer is making bike frames, it should suggest round tubes. If an engineer conceptualizing a bridge, AI should suggest cables and decking.
AI could easily reduce design-work drudgery — especially that related to the selection of standard parts, shapes, and components, and common operations. For example, when it’s found the right bracket for a given application, it might use the knowledge that at a given company such a bracket is bolted down. From here, AI could determine the size and number of screws and add the right-sized holes to the frame where needed.
3. AI should have a NATURAL INTERFACE.
Engineers can’t talk to each other without sketching; it’s a natural mode of communication. So, a natural CAD software interface would hand sketching and immediately make lines straight, snap to angles, make exact circles. Engineers should also be able to enter their concepts via sketches on a tablet. Strangely, only one CAD system allows use of a tablet — Shapr3D. Watch Shapr3D being used once and it immediately becomes clear how contrived keyboards and mice are for input and how CAD is far behind the much more natural interface of our smart phones and tablets. Notetaking software (including Apple’s Notes) can recognize and convert handwriting into text and irregular shapes into regular shapes. Hopefully the popularity of GPTs with their “natural language interfaces” and ability to understand badly formed sentences, questions, and typos will soon find its graphic equivalent in CAD.
4. AI should OPTIMIZE
CAD vendors try to pass off topology optimization and an early form of AI, but let’s stick to a definition of AI requires it to based on machine learning.
Let’s take the simple bicycle frame as an example of a design we are to optimize. One way to optimize the design, and perhaps the simplest, would be position the tubes in a shape that would use the least length of tube and therefore have the least weight. But topology optimization takes a different approach. And fails.
The challenge to optimize a bike frame is now over five years old. One company, without officially taking up our challenge, tried. But they used generative design with topology optimization, not AI. Topology optimization generates odd, random shapes until one of them satisfies initial criteria. In the case of the bike frame shown, the results were published but not the criteria, giving the impression that the generative modeling had simply run out of time without meeting the criteria: a frame superior to my bike, a diamond shape made of various lengths of tube. The absurd shapes that resulted from the topology optimization even failed as art, the last refuge of failed shapes.
If CAD companies would have let go the conceit that their free-spirited random shape generator would, all by itself, given enough time, come up with better shapes than creatively stunted engineer, had they accepted that the round tube was indeed a shape worth considering, a shape that is near optimal for torsion, as good as any for compression and tension and started with round tubes, they would have nailed the bike challenge. A simple geometry-optimizing program, an add-on to a structural simulation program, was all that was needed. Structural programs have elements specifically for tube sections. Each tube section can be one element, making the model as small as nine elements leaving out the cranks, pedals, handlebars and stem. All that is left for the program to do is vary the position of the nodes in the vertical plane until the total length of tubes is minimized.
The problem is so trivial that the reader might have already solved it with human intelligence. Need another hint? Look at the shapes upon which compact road bike frames are converging.
Ignore AI at your own peril
The message for design and engineering software vendors? Don’t try to have AI design the whole of anything. Just have AI do parts.
Keep working at incorporating AI. Engineers may not have embraced previous attempts at design optimization. It must be frustrating creating “solutions” and seeing little or no adoption. We’re sorry we’re such a tough audience. Don’t be insulted or defensive … just please, please, please look to AI to give us the help we need. There are innumerable opportunities to implement AI in meaningful ways. Don’t just check the AI box by implementing some low-value capability we never asked for or needed.
More advice for design software vendors: Don’t shrug off AI as a passing fad and think that your CAD program, having withstood decades of use, is mature, robust and most importantly, easy to use. CAD is easy to use only for those who know how to use it. For those trying to learn CAD, those that use it infrequently, it is an exercise in pure frustration.
I recently asked an executive at a CAD company why they couldn’t have a more natural interface, like ChatGPT does for text queries, to make holes for fasteners, a task I consider drudgery. I got the following response.
“By the time you create a prompt of where to put the hole, how big to make it, how many holes, and so on, the prompt is so long you might as well have made the hole themselves.”
They had, of course, had missed the point. The CAD interface should be natural — not the arcane interface that CAD programs have.
To ignore AI, or hope it goes away without disrupting the sacred code base,
venerated by the passage of decades, would be a huge mistake and extremely irresponsible. The established design software vendors need to ascertain what engineers need
now and what is possible to achieve, even if such needs are unspoken. Creating a design program
with only programming and geometry skills may have worked once, such as when AutoCAD was borne of developers without the involvement of a single architect.
That was the
2nd CAD revolution. The
next CAD revolution will be based on AI. It will undoubtedly employ principles
outlined here, namely: recognition, suggestion and a natural interface. The Big
Four (Autodesk, Dassault Systèmes, PTC, Siemens) and AEC vendors (Bentley Systems, Nemetschek, Trimble) have the resources to jump on AI and retool their software
to revitalize CAD and become the ChatGPTs of design. Whether they will remains to
be seen. Or will an startup, an upstart from Hungary, for example, seize the
opportunity AI offers and become the next Autodesk, the next PTC?
Whatever the outcome, AI will be CAD born again, this time for true computer aided design.