Big models that take a while to run

JohnM

We generally try to keep models "reasonable" in size, meaning 50-250K elements, generally less than 1M DOF.
Run times can be minutes to hours (not usually days). Lately we have had bigger models that look like below, and even using the latest PASTIX, a fast server with lots of memory and cores, I'm starting to get into the "days" category.

Any advice on solvers, hardware, methods, anything?

disla

Hi JohnM,

One variable that often goes unnoticed and I usually forget to set is "Generate solution at XXX time steps". If you are not particularly interested in the nonlinear process evolution but the final result, you can save a lot of disk writing (time) specially if there are many variables requested.

Sergio

When I was in the automotive industrie modeling rubber parts, we pay special attention to meshing (and bc) to keep the nodel count low and be able to run our no lineal problems in a razonable time. We use manual mapped mesh, hexa meshing, and control very well the quantity of elements in each volume. Obviously we simplify our models in CAD, use simmetry and all tricks availables to reduce the model size. I remember well one of my mentor's advices, limit your model to the less needed to get the required answer.

JohnM

@disla, good suggestion, we do this.
@sergio, good suggestions, we do these.
Thanks for at least reminding me this is not a trivial problem

Victor

If it's because of a lot of iterations, making the nonlinearities more gentle can save a lot, like lower contact stiffness, higher tangent modulus on plastic materials, and "loose" parts constrained a bit harder with elastic supports.

JohnM

@victor, good suggestions, we do these! The model has rubber, I'll check the contact penetration on the last run, maybe we can back off some more.

Sergio

For rubber with contact we use first order hexa elements, there you could reduce a lot the node count. We predict very accurate the load/deflection curves of the rubber parts.

JohnM

@sergio I know this trick, but also be careful, the hex8 tend to be stiff, particularly in bending. I find hex8 good for o-rings but less useful for a bellows. There is a SimCommons video on this, but I just repeated something similar to remind me of the differences, see below.

JohnM

@disla, @sergio, @victor - thanks for all of the input, here's update for what it's worth. We kept write frequency low, element count as low as possible, lowest reasonable contact stiffnesses, elastic support to keep things from "jittering". 14 hours. These models seem to just want to consume CPU and memory, so my approach will be to use more servers

Sergio

@JohnM we use first order hexas for thick rubber parts (engine mounts, bushes), and our predictions in terms of stiffness and stress (for durability) were very accurate, as we test our parts after desinging. About thin parts as bellows, even if we made it, we pay very low attention to stiffness, so I cannot made advices. Are you taking measurements of such low stiffnes parts?

Would be very interestintg to see such a model!

JohnM

@Sergio gaskets, bushings and o-rings can do okay with hex8 in compression, but for a membrane like a balloon or with mixed bending, we know that they are too stiff. Some of the examples are "really cool", but I'm sure I will never be able to share

Victor

@JohnM, this probably won't help with your big models, but going up to 5+ layers thick of hex8 can get reasonable in bending, at least for elastic behavior. The idea is to make sure individual elements don't carry any significant bending load to avoid their incorrect bending stiffness.

JohnM

The figure I shared bears out what you are saying - the cantilever with 4 layers is "almost" good enough.