MIT researchers have debuted a tool that automatically generates products–and analyzes them in detail–on your behalf.
Take these two task lamps. They each have three heads, bent and placed in very different ways. So which has the better stability? It’s a trick question. They’re equally stable–and that was discovered by an algorithm, which designed them both.
MIT researchers, in conjunction with Columbia University, have unveiled a new tool for designers who work with computer-aided drafting software. Building on previous work over the past year, their technique can optimize a design for any object, like a lamp or boat or wrench, for all sorts of metrics like mass, drag, and stress tolerance. And then it can create dozens of designs of that object, each tuned to different optimal efficiencies.
In other words, it removes iteration from the design process–and it could be applied to the design and engineering of consumer goods and industrial parts, replacing some of the human guesswork of product design and augmenting the intuition of designers themselves.
“A fundamental limitation of typical design optimization techniques is that they require a single objective function for evaluating performance. In most applications, however, multiple criteria are used to evaluate the quality of a design,” the paper explains. “Structures must be stable and lightweight. Vehicles must be aerodynamic, durable, and inexpensive to produce. In most cases, the performance objectives are not only multiple but also conflicting: improving a design on one axis often decreases its quality on another axis. In reality, designers and engineers navigate a complex landscape of compromises, generating objects that perhaps do not optimize any single quality measure but rather are considered optimal under a given performance trade-off.”
Take the humble wrench. Why is it shaped the way it is, with that bulbous head, and a skinny neck? This is exactly what MIT’s system can answer as it generates alternatives to the wrench we all know by reshaping the design to optimize for a myriad of factors at once. Through its generated CAD models, we see that a wrench with the most torque for force would actually have a very elongated and skinny design. But there’s a catch: It couldn’t take much stress. A short, fat wrench could take a lot of stress, but it couldn’t generate much torque. And in fact, right in the balance of all the most important factors–weight, torque, and stress–we see the wrench design we all know.
In other words, this machine logic can do in seconds what people perfected over centuries, arriving at the exact same conclusion–which is validating and humbling at the same time.
But of course, we know how to make a wrench, and we’ve known for a long time. What’s so promising about MIT’s research is that it can work on virtually any CAD model you throw at it, and for the most part, do so within existing workflows. That means it could help designers optimize their existing processes–and, crucially, deconstruct what works and what doesn’t, sooner.
Right now, all these design alternatives are spit out through complex graphs which are hardly navigable to the average person, and the software also isn’t available in any sort of downloadable tool for you to run. (The video you see above is from a different project from last year.) The researchers recognize this–that while they’ve created an AI that might replace a team of designers, getting it to work well for those designers is another challenge altogether: “One important consideration comes from the human-computer interaction…what is the best way to display it to an engineer who must digest the space of candidate designs?” they ask in the paper’s conclusion.
Could we become overly reliant on a single system with a single approach to design optimization, and perhaps lose some of our own creativity and ingenuity in the process? Indeed, how exactly AIs work with designers of tomorrow is the billion dollar question behind the future of design–let alone the world’s next great wrench.