[aspect-devel] problem with ASPECT - the iterative Stokes solver did not converge

[Austermann, Jacqueline]

Hi Petar,

I recently did the same run (Steinberger material model on a 3D spherical shell) and ran into similar problems as you. To all the helpful things Rene mentioned I want to add one (maybe obvious) point, which is the option to set the Linear solver tolerance up. In the end you probably want to set it down again, but for just doing test runs and checking whether e.g. your material model is input correctly if helped me a lot. I also found that Steinberger&Calderwood has a complexity that went beyond what I wanted to use and simplifying some of the parameters (not viscosity) sped up the calculation, too.

Cheers,
Jacky




On Jul 28, 2014, at 2:19 AM, Rene Gassmoeller <rengas at gfz-potsdam.de> wrote:

> Hi Petar,
> from a first glance the parameters you used for your model look good,
> but there is one thing you should test. In your parameter file you did
> not set the Adiabatic surface temperature, because it is not needed for
> the simple material model, and it defaults to 0 (@Wolfgang: We should
> maybe change this to a reasonable value like 1600? Many of the more
> complex plugins rely on a good adiabatic profile). The Steinberger
> material model, and also your harmonic perturbation initial condition
> rely on the initially computed adiabatic profile to compute their
> properties, therefore this might mess up you calculation. I would guess
> this is already the problem.
> 
> However there are a few other things that you might try to find the
> reason for the convergence problems if the above did not help. First you
> could test the convergence in 2D, in your model setup that is as simple
> as setting Dimension = 2. And you could try a higher resolution in 2D.
> Both options should show significantly better convergence, so you have
> at least a starting point.
> 
> Then you could test a temperature dependency in the simple material
> model (e.g. set Thermal viscosity exponent to something like 10). If
> this turns out to be a problem then your resolution might be too low to
> resolve the viscosity changes you are prescribing.
> 
> If it turns out to be a resolution issue, try setting the global
> refinement in the 3D model to 5 and the adaptive refinement to 0. Aspect
> will still coarsen and adaptively refine again later on, just the
> initial convergence should be much better.
> 
> You mentioned that you tested a constant viscosity profile with the
> Steinberger material model, but in the parameter file it looks like you
> only modified the radial dependence? Did you also set the temperature
> dependence to 0? You could do that by setting 'Lateral viscosity file
> name' to our test file in
> data/material-model/steinberger/test-steinberger-compressible/test-viscosity-prefactor.txt.
> With a constant radial and lateral viscosity the convergence should be
> extremely fast.
> 
> Let us know if you have any further questions.
> 
> In fact I have a question for you. We are constantly looking to improve
> the available options in Aspect and include new models. Could you ask
> Alessandro, whether he would be willing to share the viscosity profile
> you used with the general Aspect users? In that case we could include it
> in the official version as an alternative to the existing one by
> Steinberger & Calderwood.
> 
> Best,
> Rene
> 
> 
> 
> On 07/25/2014 01:14 PM, Wolfgang Bangerth wrote:
>> 
>> Petar,
>> let me pull this discussion to the mailing list since the authors of the
>> Steinberger model are there.
>> 
>> 
>>> I have experienced a problem running ASPECT. It seems that the iterative
>>> Stokes solver did not converge after three days of the job running on 48
>>> CPUs - please see attached log files (/aspect.e2477653, aspect.o2477653
>>> & log.txt/).
>> 
>> This is almost certainly an unrelated observation, but with as few
>> unknowns as you have, you should first run small models using only a
>> single processor when you try things out.
>> 
>> My rule of thumb is that you can run models with a few 100,000 unknowns
>> on a single processor. For parallel computations, you typically don't
>> gain anything if you have fewer than 50,000 unknowns per processor --
>> i.e., in your case, 4 or 5 processors is about the optimum.
>> 
>>> However, my goal is to use the depth-dependent (i.e., radial) viscosity
>>> profile (Mitrovica & Forte, 2004; please see attached file:
>>> /mitro-forte-2004-radial-visc-new.txt/) jointly with lateral variations
>>> in viscosity (see below /shell_simple_3d_new.prm/). For example, to use
>>> something similar to the Steinberger model:
>>> subsection Material model
>>>     set Model name = Steinberger
>>>     subsection Steinberger model
>>>         set Bilinear interpolation = true
>>>         set Compressible = true
>> 
>> What happens if you set this to false?
>> 
>> I don't know the parameters to the Steinberger model, so can't comment
>> on their validity. I hope someone else on the mailing list may be able
>> to provide more feedback.
>> 
>> Best
>>  Wolfgang
> _______________________________________________
> Aspect-devel mailing list
> Aspect-devel at geodynamics.org
> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/aspect-devel