Engineers have been trying to get AI to come up with good designs but to no avail. The designs AI produces are, for the most part, unusable. They are often so grotesque that they can’t be manufactured—unless they are produced with a 3D printer. But that would entail using one underperforming technology to bail out another. We might as well tie two rocks together and hope they float.
I am taking the liberty of grouping generative design under AI. Not because of the superheated interest in AI currently, brought about by ChatGPT, but because generative design fits the spirit of AI, if not the literal definition. Generative design is, after all, a computer attempting to provide an intelligent design.
Even if generative design parts are manufactured, they cannot be used for a number of reasons, including that they may satisfy one type of load but no others. For example, a stringy-looking shape that works with all parts in tension may not survive if the loads are reversed. A long thin member, like a spoke, will collapse with a small compressive load, buckling if its ends are fixed. That is an instability that is difficult to analyze for. Engineers have learned to avoid buckling from experience and testing. It is knowledge distilled through the years and kept in our heads and handbooks. Call it HI, or human intuition.
Break It Up
Human intuition can range from simple intuition, or common sense, to an analytical ability acquired from the formal study of how shapes and materials fail. One can learn how things break by breaking them, or less destructively, by studying failure theory based on mathematical principles of elasticity and field theory—all supplied in generous portions by engineering schools.
But in some ways, the math is so ponderous that it gets in the way of a true understanding, and only in the simplest of shapes is math even adequate for the task. Finite element modeling, which can break up an arbitrary shape into simple shapes allowed complex shapes to be analyzed. Finite element modeling was a perfect application for the computer, which made short work of the analysis of hundreds, even thousands of the simple shapes a part had been chopped into.
Finite element analysis (FEA) and its fluid analysis counterpart, computation fluid dynamics (CFD), are the methods used to validate iterative steps in topology optimization. For the sake of argument, let us stretch the category of AI to include analysis software—even though it’s only because a human who could solve hundreds or thousands of equations in short order would be considered a genius.
What’s Wrong with Analysis Software?
Analysis software has these major problems:
- It is not accessible to everyone.
- It is not always accurate.
- It takes too long.
It is not automated.
Accessibility: Few can afford it and few can use it. Never mind the democratization of simulation attempted by the same companies that are selling generative design. To really understand and effectively use analysis software, you should have training (months), education (an advanced degree in engineering), and experience (years).
Accuracy: The results of analyses need to be validated somehow because they can be way off. Not only does breaking up continuous shapes into elements and cells introduce error, but it’s also super easy to provide bad input. In addition, the mesh may have been too coarse, the wrong failure theory applied, or the wrong turbulence model used … the list is much longer, but you get the point. There are a lot of ways to get the wrong answer if you don’t know what you are doing.
Time: Accuracy is proportional to the number of cells, but increasing the number means taking more time. Computers getting faster only resulted in models with more elements. Though there is near real-time FEA available now, it is limited in application. Simulation time expands to whatever time you have, whether it is a coffee break, the afternoon, or, in the case of complex systems, overnight. The point is: you are going to have to wait.
Lack of Automation: Analysis results means that your software is done. But your work is just beginning. Results with stresses over the material limit will make you thicken a part, for example. Then again. This makes analysis a manual rather than an automated method.
Analysis could be more useful if it were used with optimization routines that would add or remove material based on the stress on the element. It is simple enough to add more elements where stress is too much and remove elements where there is little or no stress.
It would be nice if optimization worked that way—if it started from your part design and tried to improve it. But such is the conceit of optimization software, or should I say, optimization software vendors—that they insist on starting from scratch, with a blank screen. As if to say, thanks but no thanks. Just tell me what your requirements are. I’ll figure out the shape. I got this.
So off goes every generative design program, speeding toward a solution—one that satisfies your requirements. It’s the fastest retriever imaginable. It’s so quick, it will try again. It will use a different route and retrieve another solution. If you have left it running overnight like one of your most complex simulations, the generative design program will greet you in the morning with hundreds of solutions.
You thumb through a few … but you get that sinking feeling. You can’t use any of them.
To be continued….