Failure to converge with scaled up model

DaveStuppleDaveStupple
Hi, I've got a quite large model with (many quad elements, all hex20 except a the heat load). Thermal steady state was solving fine with CCX until I had to refine part of the mesh to make the internal 1heat load more concentrated. I used Mecway's ref tool to refine a small region of mesh and ended up with a few point elements pyramids and the like. On hitting solve, CCX immediately threw an error (non-zero jacobean?). Last time I had this error, Victor came up with a workaround of scaling up the dimensions of the whole model. I did this (x1000) and left all the properties/loads as they were before. The model solved fine with a very uniform temperature, predictably. When I scaled up the thermal conductivities and heat transfer coefficient to match the size, the model failed to converge, dropping to around 10^-13 before shooting up again. This happened every 5 to & iterations, so CCX would start again with a smaller increment. A much simpler model that I tested the scaled properties on seemed to converge between 10^-10 and 10^-11. I thought maybe the heat loading or flux was too high so I scaled the original model to x10^6 in the same way. This too failed to converge. I had a vague memory that CCX had a problem with pyramids, so I am replacing them with tied contact for my next attempt. I am not confident this will improve things, as the other model (before I scaled the constraints) solved OK.

Any ideas? Scale to 10^9? The smallest corner node distance is already 250 mm!

Thanks for reading.

Dave

Comments

  • DaveStuppleDaveStupple
    edited January 23
    Line 1 should read 'except a few wedges away from the heat load' and line 8, '5 to 7'.
  • JohnMJohnM
    The pyramids are occasionally cursed. I have had problems with this using the local refine, mostly on structural stuff. With this being a thermal and only 1DOF/node, can you get away with refining the whole model?
  • DaveStuppleDaveStupple
    Thanks JohnM, appreciated. I could certainly give it a go. It will be ungainly. First I'll try all hex and tied contacts. I'm nearly in a position to run this.
  • VictorVictor
    Since you've already scaled it up enough to avoid the negative Jacobian error, I don't think there's any value in scaling it up beyond that. It might push something over the limit in the other direction.

    Pyramids are mostly bad for boundary conditions and accuracy but you could delete them to see if it fixes it.

    You can change to linear elements for faster iterations in debugging it.

    You can also try deleting big chunks of the model until it works. You can usually pretty quickly zero in on what's causing the problem that way.

Sign In or Register to comment.

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Sign In Register