The design software industry, eager to push technology with such potential, acquired at great cost, sees the rejection of generative design as evidence of engineers who are stuck in their ways, content to work with familiar but outdated tools, in the dark and unable to see the light and realize the potential of a game-changing technology. Engineers, on the other hand, say they never asked for generative design—at least not in so many words. For this, software makers quote Henry Ford:

“If we’d asked people what they wanted, they would have said faster horses.”

As if they have the prescience of Henry Ford. As if they somehow knew what engineers wanted—despite what they said they wanted. As if generative designed parts should look like they do.

“It’s not weird; it’s wonderful,” they would say of their generative design parts.  “I know you better than you know yourself," they say, as would a parent scolding a child. The part may not look like you expected. That’s because generative design is a free spirit—and you are trapped in a box of your own making.

Like 3D printing, another technology desperate for engineering acceptance, generative design sees its “solutions” as perfect. One such solution was a generatively designed bracket. The odd-looking part was discussed as a modeling experiment by Kevin Quinn, GM’s director of Additive Design and Manufacturing, but with no promise of mass production. It was obviously fragile and relied on 3D printing for its manufacture, making it unmanufacturable at the quantity required. It may have withstood crash test loads, but reverse loading would have splintered it. Yet, the part was to appear in every publication (even ours) and almost everywhere lauded as a victory for generative design if the saint of lightweighting, a pressing automotive industry priority.

Now more than ever, engineers find themselves leaning into hurricane winds of technology and a software industry that promised us solutions. We are trained to accept technology, to bend it to our will, to improve products we design, but the insistence that software has found a solution to our design problems with generative design puts us in an awkward thanks-but-no-thanks position. We find ourselves in what Gartner refers to as “the trough of disillusionment.”

That is a shame for a technology that, if it were to work and evolve, could be the “aided” in computer-aided design. (For the sake of argument, let’s say that computer-aided design as it exists now is no more than an accurate way to represent a design that an engineer or designer has a fuzzy picture of in their heads).

Generative design should take a part and make it better. That’s practically a no-brainer. Whittle away material that will not be stressed, for example, until you’re left with only with the material necessary.

But then we want more. We want to be able to manufacture the generative designed part the same way we manufactured the original part. If it was NC machined, the generative designed part should be NC machinable. Can we not have every one of the so-called solutions only producible with 3D printing? And don’t scold us for not accepting 3D printing. To engineers, that is another technology that has overpromised and under-delivered.

If our part were designed with tubes, I-beams or cables, would you accept our reasons for doing so? You might have thought that we had only hammers and everything looked like a nail. But we designed with standard shapes because those shapes work. They are in of themselves optimized shapes. You simply cannot do any better.

For example, the round tube, with material around a central axis, is perfect in torsion. The I-beam, its flanges farthest from the center, has maximum material to resist bending—its primary purpose. We’ve had centuries and generations to optimize these shapes. You could spin your wheels trying to optimize them further—or you could just use them as they are.

Is it any wonder that generative design put to the task of coming up with a bridge to span the Golden Gate produced what can only be described as a monstrosity due to its inability to use cables—the key structural element of a suspension bridge. A cable is the most efficient way to hang a load. Even AI should know that.

Is it any wonder that no generative design vendor has accepted our challenge to build a better road bicycle frame from three years ago? The road bike frame, with tubular construction, uses a minimum amount of material and is rigid against forces and moments. The recreational bike rider may only be concerned with up and down forces, but a competitive rider cannot tolerate the twisting from the massive amount of torque generated from high pedal forces that is resisted by the handlebars. The generative designs of bicycle frames have consisted only at hacks at mountain various that focus on radical-looking shapes but conveniently neglect to mention weight savings. The road bike racing community, for whom weight is a religion, have little use for generative design—except for an odd part here and there, valued more for its artistry than its performance—such as the part below.

How much trouble would it be to add some of what we know—our insight—to generative design? After all, that is another technology the software industry is fond of pushing. Watching a topology optimization take shape can be about as painful as watching a roomful of monkeys banging randomly on a keyboard and hoping to write a Shakespeare play. If, by some miracle, they form “What light through yonder window breaks?” our only hope of the right answer would be to type it ourselves. Similarly, an optimization routine starts creating a stringy shape. Bam! Let’s make it a cable and move on. A smooth shape is forming? Jump ahead and make it a flat surface. See a gap forming? Make it a machinable slot. Know a frame will undergo torsion? Stop the madness and use a round tube. (The shapes made with already optimized elements can still be optimized by adjusting angles and lengths.)

The inclusion of AI is what is strangely absent in generative design to this day. We are reminded of a recent conference (pre-pandemic, of course) in which we saw a software vendor go around a generative designed shape, replacing it bit by bit with standard shape elements—a round rod here, a smooth surface there. Really? We should have to do that?

Classical optimization techniques are a separate technology. Like CAD and CAE, they are based on mathematics. Unlike CAD, they have their own language. Optimization borrows language and nomenclature from calculus (optimum, dy/dx = 0, etc.) and adds some of its own. While optimization can be applied to any phenomenon, its application to 3D shapes is most relevant to this discussion. Each iteration of a shape is validated with a numerical technique. For structural shapes, the validation is done with finite element analysis (FEA). For fluid flow optimization, the validation is done with computational fluid dynamics (CFD). Therefore, the application of generative design uses the language of simulation, with terminology like boundary conditions, degrees of freedom, forces and moments. It’s a language foreign to designers and forgotten by the typical product design engineer that runs counter to the democratization of generative design.

The best technology is one that just works, requires little learning, and may not even need an introduction. Think of AI implementations by Google, delivered to our delight, with no fanfare—not even an announcement. Here was Google correcting our spelling, answering our questions, even completing our thoughts and translating languages. Scholars skilled in adapting works from one language to another were startled to find Google equally skilled. Google held no press conference, issued no press release, or even blogged about the wondrous feat of AI. It just worked. And it required no learning.

By contrast, IBM trumpeted its AI technology, Watson, after digesting the sum of human knowledge, easily beating Jeopardy! champion Ken Jennings. But when it came to health care, Watson bombed at the very task it was most heavily promoted for: helping doctors diagnose and cure cancer, according to the Wall Street Journal.

The point is quick success and acceptance will be had with technology that seamlessly integrates into how people already do things and provides delight and a happy surprise. As opposed to retraining, asking users to do things in a whole new way with a new, complicated application that requires them to learn a new language or terminology.

The generative design application that can almost be guaranteed acceptance is one that works seamlessly inside a user’s chosen design software. It could be working in the background, using idle CPU cycles to suggest different shapes as you are sketching or shaping a product as Google does, guessing almost without fail, what you intended to type, rather than what you actually did. AI could certainly guess that an exact shape is intended—a perfect circle, for example—if a circle is sloppily sketched. Some CAD applications can already do this. Harder yet, but still within the realm of possibility, is to guess a mathematically optimized shape—or even a shape that exists in a standard part library, or to match a shape with a similar or exact duplicate of the company’s existing products. Reinventing the wheel is a common waste of time and effort in large, far-flung enterprises that have developed silos of information. (This borrows on shape search technology, not optimization, but is included here as it is another technology, like optimization, that should be integrated into CAD.)

When you are done with a design—or simply have run out of time—wouldn’t it be nice to have generative design so advanced that you could push an “optimize this” button and let AI and generative design have a crack at it? But before you push that button, you must have the confidence that generative design will stay in its lane. In other words, the refinement generative design provides will still be manufacturable by methods you intended and doesn’t look like it’s made from twigs. In this scenario, you would consider only designs that are truly improvements, not artistic flights of fancy.

Repeated use of the “optimize this” button resulting only in designs that offend an engineer’s sensibilities could cease to be used—an unhappy outcome in the short term but important as a true indicator of the technology’s success for CAD vendors.

Products fail not from causes we consider but from those we neglect to contemplate. The odd-looking, stringy brackets mentioned that are sure to fail under compressive load are frequent examples. A thin-as-paper composite bicycle frame may be fine for zipping along a smooth level road, but it will collapse on impact. In 1940, the Tacoma Narrows Bridge failed because engineers did not consider resonance resulting from wind. Galloping Gertie, as It came to be known, has been shown to every engineering student since. You’d think we would never miss another catastrophic load case. Yet, here we are. We have 3D-printed parts in metal that not only introduce additional risks (voids, prestress, distortion, annealing) but are also obviously subject to a single load case.

The generative design vendors may want to go back to the lab and resist the temptation of trotting out half-baked solutions. That is clearly not helping. The technology, which relies on bone growth algorithms, massive cloud resources, optimizations that are not, and so on, all make for a technology that is not ready for our practical use.

In the meantime, we promise we will keep playing with generative design. After all, it is pretty cool. And we can imagine ourselves creative and artsy for all the generative designs produced that we can’t put into production.