Overcoming convergence problems

Anyone have an opinion about the best way to overcome convergence problems?  What do you find to be most effective - refining the mesh, or reducing the time-steps (or something else)?

Is there any method for trying to resolve convergence issues that is more efficient than simply experimenting?

Convergence problems are very problem specific, but often arise due to poorly formed elements.  Run the mesh checker (in classic); if you have more than 10% "red" elements, you need to smooth or reformulate the mesh.  Sometimes, the autostepping can get bogged down working on a small area in the model.  You can fall back to the traditional user-specified stepping in which you grow the steps using a multiplier of 1.1 to 1.3 and start from a very small value, like 1e-6.  Also, do not let your steps grow too large.  In transient models with a lot of stress variation, user-stepping can get very complicated.  When using user stepping, you must check the mass balance is ok and that the local balance is ok in areas of concern (because user stepping may not yield optimal results).  Experimentation is necessary. 

Pete

Thanks Pete, that was really helpful.  I have been running a big unsaturated model and have found that convergence can be a challenge using the autostepper with no maximum limit.  When I limit the step to 0.01 days, the model takes a day to run, but I get convergence.  So what you are saying fits with what I see. 

Agree with Pete. Its problem dependent and you really need to check out your grid. My experience is that quadrilateral elements work better, so if you have a choice ... The size of the elements in the Z-dir is also important. You might have to insert several numerical layers to get stability. I don't usually employ user stepping. As Pete says you need to check the results all the time for errors. Prefer to reduce the error tolerance to 10[sup]-4[/sup] or 10[sup]-5[/sup]  ( using FE/BE) much slower but more stable. A 30 year simulation with e.t.10[sup]-4[/sup] took 2 days with some oscillations but nevertheless... With e.t 10[sup]-5[/sup] no visible oscillations but took almost 4 days. I suppose that managing the errors is part of the "fun" of modeling.

cheers
Panos

AB/TR may work better in some situations (UZ problems for one)
Pete

Pete and Panos:

Those were both really helpful posts.  Thanks!