Generative Design converging early with low iterations?

[dsp39]

Hi all,

I've tackled a couple simple problems in Generative Design (and attended Autodesk's training in PDX, which was very helpful). My simple efforts were pretty straightforwardly successful.

I'm working up a more complicated generation that is modeling the support structure between a violin chin rest, shoulder rest, and cylindrical attachment point to the instrument.

I've thrown six different studies at the system and have had disappointing results for all of them! I'm consistently getting about 11 iterations and a very bulky result. Perhaps I need higher resolution synthesis? I'm on the default. I suspect this is somehow related to my starting shape. My first generation or two did not have one and gave me some more interesting results. 

Here's a screenshot of the latest study and a few of the bulky, VERY high safety factor results. I've been dropping my loading forces in the hopes of generating better results, but I wonder if (based on the 0 MPa stresses shown in the results) that's a mistake. 

 My first study, with no starting shape and similar loading, spat out much more interesting results. Bonus question, why are Outcome 1 and Outcome 2 SO different? The only difference I identify is the "Makerbot" PLA vs "Ultimaker" PLA. 

In summary:

1. How can I get more relevant results for the studies shown, above?
2. What is causing these high safety factor and high mass results? 
3. What role is the starting shape playing (if any) and how can we make it better? 
4. bonus: Why are outcomes 1 and 2 so different from study 1?

-David

[kenikd]

Hey David,

Thanks for posting in the Forum.

If I understand your question correctly, you are asking why you are seeing more organic results with no starting shape, vs bulkier results with a starting shape? Is so - the use of the starting shape is suggesting an initial design space for the technology, which is restricting the space it can "play" in. I would suggest either (a) not using a starting shape, and creating more obstacles, or (b) using a more interesting starting shape as a "suggestion" - lattices tend to work well here.

For the bonus question - I would take a look at the material Stiffness and Poisson Ratio. They will likely be different, which is why you are seeing different results. This is evident my the maximum deflections between the 2.

Hope that helps,

-Doug

[dsp39]

Doug -- thanks for the reply. 

Honing in on the starting shape...it seems that if I limit the space in which the design can play, then it would spend more time exploring 'interesting' solutions within that space. Why is it that with a starting shape the final safety factor for these studies is orders of magnitude higher than the goal? 

Related, what is causing the software to stop iterating? There's a lot of material here that could be removed!

Yeah, those two versions of PLA have slightly different properties. Typically, though, they yield very similar solutions. This is the first time one has resulted in a solution that's drastically different from the other, so I wonder if something else is going on.  

-DP

[kenikd]

Not necessarily... think of a corollary here.

Say you want to find an interesting volume to fill with a specific amount of fluid. The more you restrict the space the same amount of volume the fluid can fill, you will eventually run out of space to do anything - the volume you can fill will be the same volume you give it with a seed geometry. The more space you give the solution to play, the more interesting designs will come back.

For the second part... just look at the deflection compared to the max stress of the two parts. This is a large discrepancy. Maybe the point in the design space chosen was incompatible, but I would think the material properties are driving it. We can investigate later if needed.

[dsp39]

OK, cool. I understand how limiting the volume that can be occupied could also limit your results and their interesting geometry. What I don't understand is why the generation would converge at a safety factor >1000. Why does the software declare the problem solved after only 11 iterations and at such a high safety factor? There is still room inside the assigned starting shape to remove material. Does that make sense? 

-DP

[kenikd]

Hey DP,

I think we need to clarify what you are asking the software to do. Your request is likely "Minimize Mass" and you have input a FoS of something like 2, I assume. The FoS you input is a lower bound, it is not a target. Your request is for the software to try to find a lightweight design while keeping the FoS above 2 - in this case. 

The way the optimization algorithm works in the background is tricky, and I cannot disclose the specifics. However, it has hit some sort of "local minima" in its quest, and sees this as an acceptable solution. So, in theory it is actually solving your request, although it might be able to perturb itself and do a bit better. Something we can look into.

-Doug

[dsp39]

OK, this is starting to make sense. 

What I feel like I need to know is why the software has deemed this lightweight and how can I convince it that actually it isn't. I hope you can understand why, from my ignorant user perspective, the results are disappointing and confusing.

If this sort of situation were to arise again, what should I do? Add a keepout in a random place? Put some cutouts or holes in my starting shape? 

It's unfortunate you can't say more about what the software is doing. I think I might be able to use it more effectively if I better understood what's going on behind the curtain. Not to mention the fact that I spent 125 credits trying to make it generate something lighter and got nowhere. 

For now I've abandoned that last project and am moving on. I may return to it later.

[kenikd]

So...

I would like to point out that generative is a design space Exploration tool to provide inspiration for new designs. It is in the system's best interest to promote designs which may or may not be the "most optimal" for a specific application, because there are so many subjective modifications which could be made downstream due to other requirements. I am not arguing these types of designs you are pointing out should be considered, they are a consequence of the what the system is trying to accomplish, and we should continue to determine how to influence it properly.

If this does arise again, and you are not satisfied, I would either (a) increase the surface area of the starting shape (make it a lattice, punch holes in it, etc.) or (b) place more obstacles in the design space. As you play with it, you will begin to understand how it works, and we are continuously trying to improve it so you do not feel cheated with cloud credits. Stay tuned - I think you will be pleasantly surprised as we continue to develop the technology.

-Doug

[dsp39]

Thanks for those tips! Sounds like surface area really is key. Great -- I'll know what to try next time it seems like I'm not getting the results I expect. 

Haha, yes, I am admittedly unsatisfied with my results for this specific attempt and disappointed by my rapidly decreasing credit balance. I'll say that most of my other generative projects have gone quite well and I am happy with the technology and its implementation. That's part of what made these results confusing! 

I'm excited to see where y'all go with it all and I'll play with surface area. Thanks, Doug!

-David

[I_Forge_KC]

Here is some (light) reading that may help you wrap your head around the algorithm at play...

https://forums.autodesk.com/t5/fusion-360-design-validate/what-is-the-difference-between-generative-...

The usage of lattices (or generally high surface area to mass ratios) is really helpful in broadening the domain you're working in. This is an example of a cell phone holder I designed some time ago. Each of the shapes met the functional requirements, but the addition of various starting shapes and lattices gives new life to each run.

[evs5454]

This seems like the best thread in which to put this:

I had a generative design study whose outcomes were all merely one iteration (no change from the initially shape):

1. Under what circumstance is convergence on the 1st iteration not an error? I can’t understand why an outcome identical to the initial model would ever be a desired result.
2. What causes 1 iteration convergense ? I took the precaution listed above. ‘Perhaps to model has walls too thin?