Generative design, for readers who may be emerging from an extended sleep, is the algorithmic, iterative attempt to produce shapes given nothing but design constraints. It is all the rage in CAD circles. Its champions call it the biggest change in CAD since… CAD, itself. They insist that it is—or soon will be—the way design is done, with computers and software finally rising to their full potentials, able to design, not just model. Its critics look at GD as they would a thousand monkeys banging away on typewriters and the unlikely chance that one would produce something remotely literary.
One of these fell as is bound to write a better article than this one. Critics of generative design give monkeys a better chance of writing a Shakespeare play, a sonnet, or even a verse, than of making a useful design.
At ASSESS 2018, one presenter after another reveals a version of GD. Several GD vendors are in attendance. As a group, they form a big set of newcomers to the fledgling conference. This is only the 3rd year of the Analysis, Simulation and Systems Engineering Software Strategies, which exists to advance the cause and broader acceptance of FEA and CFD, the mainstays of simulation. GD does not present as simulation because its computational methods are out of sight. Its solution is under the hood.
With the amount of attention being paid to the newcomers, there are more than a few attendees feeling like they are old news.
Generative Challenges
Generative design and topology optimization was theory spawned in another century and, for years, it was kept in the lab, the subject of research and the stuff of theses. It took too much power, its number-crunching so intense, only universities with mainframes and supercomputers could use it. The recent advances in PCs and GPUs have allowed a few generative design shapes to emerge from the lab. Hardware is now fast enough to handle the number of iterations GD needs to run through. However, its use as a design is still questionable. So far, we have seen a few experiments, either whimsical in form like artwork or odd, stringy shapes that look like they’ve been stretched apart like taffy. They invariably come with descriptions of how much weight they save—a most impressive weight savings, in fact, like 50 percent or more. As if that compensates for looking weird.
We hear promises that GD will tame its shapes, try to make shapes more like those prismatic shapes, easily mathematically rendered and easily machined. For now, most GD-produced shapes are so irregular that they can only be represented with triangular meshes or STL files. (Autodesk Fusion 360 goes an extra step to create geometry with “T-splines.”)
Beneath the Surface
The complexity of shape increases with internal structures, such as lattices or cellular structures. A solid part can be made much lighter—and while looking the same and preserving most of its strength—by hollowing it out. But such internal structures add so many features to the 3D model that CAD programs quickly give up. It’s hard to store the irregular geometry. It’s also hard to display the detail on the screen. A small part with an internal cell structure of your femur could have as much detail as the airplane that flew you to the conference.
Nature has had much more experience, heard at ASSESS. This section of human femur shows an irregular, space-filling structure, just about impossible for CAD to model and CAE to solve. (Picture courtesy of ASBMR Bone Curriculum.)
If modeling topology-optimized and generatively-designed shapes is difficult, the simulation of GD parts with internal structures is impossible. Modeling each little leg of a lattice or each wall of a cell, either with beam elements or thin shell elements respectively, begs for compute and storage resources that are out of this world. Think of simulating the weather by modeling air molecules.
The answer at the moment for simulation for irregularities, external or internal, maybe in being able to approximate the internal structure with a somewhat equivalent solid structure. That’s tricky. There could be equivalent mechanical properties for one type of loading or environment and not for another. It is a field of study known as “multiscaling,” with its own champions suggesting they have it all figured out. They don’t. Multiscaling become difficult as the internal composition varies, as it does in nature.*
“Nature has had much more experience,” says an ASSESS attendee.
*In all fairness, one vendor (nTopolgy) does claim to be able to solve gradually varying internal structures but a demo was not available in time for this article.
Comments