Generative design issues

[arved]

Hithere, I have a generative design study that completed without convergence at iteration 52. This is a bicycle frame and the manufacturing process was unrestricted. 

Up until iteration 24 the frame started to converge nicely looking more and more like a bike and loosing weight. There were no significant stresses at iteration 24.

Iteration 24

After iteration 24, the design started to gain more and more weight in apparently unloaded areas and ended completed without convergence at 52 looking rather clunky.

Iteration 52

This does look like the algorithm going rampant somehow. We did go back and edited the model in the generative design space perhaps around iteration 24, but I was under the assumption that once submitted, the study does not update based on changing obstacles, load cases, or required geometries. If someone could confirm that that would be great. 

Otherwise I would like to know if there are pointers to the setup of this model that could lead to a better design outcome.

Thanks 

Arved

[I_Forge_KC]

I'll throw a couple things at you. Take them all with a grain of salt as I'm a power user but not a mathematician.

1) I've found that larger models like this often suffer from issues because the voxel size is a tad too large. My thinking on this is that there exists what you might consider to be a singularity within a certain voxel (infinite stress). It's not "revealed" until the surface boundary moves to its location. It doesn't show on the stress plot because it's also infinitely small (or perhaps, between voxels - the larger voxel size prevents you from seeing it). The solver then reacts in a contradictory fashion by trying to thicken around the singularity. Since the stress is infinite, the thickness would also have to be infinite to support it - hence why the model changes from thinning to thickening until the iteration timer runs out. As for solutions, try these:

• Shrink the voxel size
• Try using a starting shape

2) Models with a very high reduction in mass often need a head start. The mass burn down rate of the solver is dependant on surface area (level sets work by moving the surface boundary). If your default starting shape has a very high volume to surface area ratio, the solver will be slower to remove material. A solution to this is to use a starting shape with a high surface area. Try punching a bunch of holes in the starting shape to create surface area. You can do this by using an explicit starting shape OR you can add a bunch of wacky obstacle geometry. I prefer the former for a few reasons - 1) the solver occasionally forgets input geometry from iteration to iteration, 2) you're providing undue user input that will change the output shape (though this may be intentional), 3) if you're in control of the starting shape you can leverage exotic shapes.

You are correct that once you submit a study it is tied to the version it was saved with. If you make any changes to the model, a new study would need to be run to incorporate those changes.

[arved]

That is fantastically insightful and thorough, thank you so much. I did not know, for example that surfaces are being removed and not volumes. Is it possible given this, that generative design would ever come up with a hollow tube as the best solution for a torsion problem? Are hollow initial shapes acceptable? Questions over questions.

Meanwhile our thing is looking more like a bike.... Thanks for all.

[I_Forge_KC]

Again, grain of salt.

If you scour the internet looking for videos on Dreamcatcher (the spiritual core of generative design) you'll find videos from the research team that mention the shapes being created are based off a bone-like approach. They jokingly mention that it makes mammal pelvises out of every design. There are other shape-solvers out in the wild (e.g. spaceframe/beam solvers) but the specific one implemented here is an organic "solid" form.

During the beta and early releases I spent a ton of time trying to wrap my head around the solver quirks, including forcing it to make hollow structures, lattices, and other oddities. The specific shape-solver tends to create trusses instead of thin-wall parts. It's just in the nature of it I suppose. Here is an attempt that we printed during my Pier 9 residency showing the trusses against an array of spheres with a hollow center...

My hope was that we would get either a nice lattice or maybe a tube. I got a lattice-like form, but it is clearly parabolic struts between nodes here and they all have the organic bone-like surface shape.