[aspect-devel] Problem: Iterative advection solver does not converge

[John Naliboff]

Hi Felipe,

I would just stick with the default values as you build your model up.  Do note, though, you can still get large gradients in strain-rate even if a model is isoviscous.

Cheers,
John



> On Jun 6, 2016, at 2:01 PM, FELIPE ORELLANA ROVIROSA <f_orellana at berkeley.edu> wrote:
> 
> 
>  Hi John,
> 
>     I understand your points.
> 
>     I have tried tolerances in the range 10 ^ (-6 to -8 ).
> 
>     I am trying to combine them with max number of iterations (10,20,100), but there are many combinations, many I haven't really tried.
> 
>     My system is isoviscous 10^21 Pa.s, so I kind of expect not to have dramatic spatial variations on the strain rate..
> 
>     I am indeed planning to use temperature-dependent viscosity, and perhaps non-Newtonian stuff, in the future. Regarding that, I think I first have to understand the more basic case I am dealing with now.
> 
>     I could start just taking my unitary-dimensions model (which runs successfully) and upgrade its dimensions, but doing one-by-one might take me weeks.. So I am trying to make an educated guess on what to change, somehow a shortcut..
> 
> thanks a lot for your comments,
> 
> Felipe
> 
> On Mon, Jun 6, 2016 at 1:46 PM, John Naliboff <jbnaliboff at ucdavis.edu <mailto:jbnaliboff at ucdavis.edu>> wrote:
> Hi Felipe,
> 
> In general, it is best not to ‘force’ a solution to converge by decreasing the solver tolerances significantly. 
> 
> Personally, I never decrease the Linear solver tolerance below 1e-7 … I have used values between 1e-7 and 1e-9.
> 
> The nonlinear solver tolerance is a bit more tricky if you using highly non-linear constitutive relationships (i.e. strain-rate dependent viscosity, plasticity, etc).  In reproducing some published shear band (plasticity) models, I can only get the non-linear residual to converge down to 1e-5 or 1e-6 six if I’m lucky.
> 
> In general, convergence behavior may vary significantly from problem to problem.  If you are using a highly non-linear material model, you will need to try varying the convergence parameters to ensure your key findings are not highly dependent on these values (they may well be).
> 
> Does this help? 
> 
> Cheers,
> John
> 
> *************************************************
> Assistant Project Scientist, CIG
> Earth & Planetary Sciences Dept., UC Davis
> 
> 
> 
> 
> 
>> On Jun 6, 2016, at 1:25 PM, FELIPE ORELLANA ROVIROSA <f_orellana at berkeley.edu <mailto:f_orellana at berkeley.edu>> wrote:
>> 
>> 
>>   Hi guys, 
>> 
>>    I still have a quick question:
>> 
>>    In your experience, what should be the typical values for the solver parameters
>> 
>>      Linear solver tolerance             ..not far from the default 1e-7 ?
>> 
>>      Max nonlinear iterations            (I have used 10,20, 100)?
>> 
>> Felipe
>> 
>> On Mon, Jun 6, 2016 at 1:11 PM, Wolfgang Bangerth <bangerth at tamu.edu <mailto:bangerth at tamu.edu>> wrote:
>> 
>> However, I think the definition of the residual may be a bit different than
>> what I’m used to:
>> r = || v_i - v_i-1  ||  / || v_i  ||
>> where v_i is the current velocities and v_i-1 are the previous values.
>> 
>> The "residual" (literally: "what is left") in iterative solver methods is typically the quantity
>>   r_i = A x_i - b
>> (or the norm of this vector), where x_i is the solution vector in the i-th iteration. It can be "large" because it carries physical units. As a consequence, its numerical size is meaningless unless compared to, say, the norm of the right hand side itself.
>> 
>> For example, we typically terminate iterative solver if the condition
>>   || r_i ||    <=    10^-6  || b ||
>> is satisfied. What is relevant is not whether the numeric value on the left is large or small, but whether it is large or small compared to ||b||.
>> 
>> Does that help?
>> Cheers
>>  W.
>> 
>> 
>> -- 
>> ------------------------------------------------------------------------
>> Wolfgang Bangerth               email:            bangerth at math.tamu.edu <mailto:bangerth at math.tamu.edu>
>>                                 www: http://www.math.tamu.edu/~bangerth/ <http://www.math.tamu.edu/~bangerth/>
>> 
>> _______________________________________________
>> Aspect-devel mailing list
>> Aspect-devel at geodynamics.org <mailto:Aspect-devel at geodynamics.org>
>> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/aspect-devel <http://lists.geodynamics.org/cgi-bin/mailman/listinfo/aspect-devel>
>> _______________________________________________
>> Aspect-devel mailing list
>> Aspect-devel at geodynamics.org <mailto:Aspect-devel at geodynamics.org>
>> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/aspect-devel <http://lists.geodynamics.org/cgi-bin/mailman/listinfo/aspect-devel>
> 
> _______________________________________________
> Aspect-devel mailing list
> Aspect-devel at geodynamics.org <mailto:Aspect-devel at geodynamics.org>
> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/aspect-devel <http://lists.geodynamics.org/cgi-bin/mailman/listinfo/aspect-devel>
> 
> _______________________________________________
> Aspect-devel mailing list
> Aspect-devel at geodynamics.org
> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/aspect-devel

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.geodynamics.org/pipermail/aspect-devel/attachments/20160606/7ba30e04/attachment-0001.html>