[aspect-devel] Far different velocity magnitudes & timestep sizes of the same Ra

[Scott King]

We’ve run with everything from 10^{-2} to 10^{-7} and seen no difference.   It was on Grant’s AGU poster.
You have no idea how many times this problem has been run on our cluster….   That said, Shangxin I suggest
you rerun it just to check.   I’m not sure Grant did this after the Bousinessq formulation was released.


> On Apr 20, 2017, at 4:47 PM, Timo Heister <heister at clemson.edu> wrote:
> 
> Shangxin,
> 
>> set Linear solver tolerance                = 1e-3
> 
> Is there a reason for choosing such a tolerance? I would have no
> confidence that the solution is accurate. Can you re-run with
> something like 1e-8.
> 
> That said, I think we likely still have a problem with equation
> scaling if the gravity is not the same order of magnitude as the
> velocity.
> 
> 
> On Thu, Apr 20, 2017 at 3:01 PM, Shangxin Liu <sxliu at vt.edu <mailto:sxliu at vt.edu>> wrote:
>> Hi Timo and John,
>> 
>> Before making further comments, there is a typo in my first email. I just
>> checked my prm files for these three cases. The CFL number of them is all
>> 0.5 instead of 1. And I use 3 global refinement, 0 adaptive refinement, with
>> fixed mesh.
>> 
>> Yes. I can further try decreasing CFL number or increasing the grid
>> resolution. But the reason why we use 3 global refinement, i.e., 49152 cells
>> globally is that Zhong et al., 2008 paper uses 12*32^3 (393216) cells
>> globally, which is equivalent to global refinement 4 in ASPECT. However,
>> since ASPECT uses quadratic element while CitcomS uses linear element, we
>> suppose the lower global refinement 3 of ASPECT is equivalent to higher
>> global refinement 4 in CitcomS.
>> 
>> John, the reason why we modify alpha and gravity to fix Ra is that
>> previously before the true Boussinesq approximation was merged into ASPECT,
>> we use a very small alpha to mimic constant reference density in energy
>> equation. Now we further try these series with the new ASPECT with
>> Boussinesq approximation formulation and found this confused issue.
>> 
>> Timo, I'm using free slip boundary condition for velocity on both top and
>> bottom boundary.
>> 
>> The attachment is the prm file for my first case that uses 7000 gravity and
>> 1.0 alpha. You can simply change gravity to 70000/700000 and alpha to
>> 0.1/0.01 to get case2/case3. Let's see whether we can get the same tilmestep
>> size and velocity magnitudes of the three cases.
>> 
>> Best,
>> Shangxin
>> 
>> 
>> 
>> On Thu, Apr 20, 2017 at 1:09 PM, Timo Heister <timo.heister at gmail.com>
>> wrote:
>>> 
>>> Shangxin,
>>> 
>>> can you post your .prm? Are you using free slip boundary conditions
>>> for the velocity?
>>> 
>>> On Wed, Apr 19, 2017 at 10:17 PM, Shangxin Liu <sxliu at vt.edu> wrote:
>>>> Hi;
>>>> 
>>>> Recently I'm trying the new Boussinesq approximation formulation of
>>>> ASPECT
>>>> to test the 3D spherical shell time-dependent convection. I tried
>>>> running
>>>> the non-dimensional cases and set the start time to 0 and end time to 1.
>>>> Through a series of tests, I found that the same Ra but different
>>>> gravity
>>>> and alpha have far different results of velocity magnitude and the
>>>> tilmestep
>>>> size. For example, I fixed the Ra to 7000 and run three cases:
>>>> 
>>>> Case1: gravity 7000, alpha 1
>>>> Case2: gravity 70000, alpha 0.1
>>>> Case3:  gravity 700000, alpha 0.01
>>>> (All other parameters of these three cases in prm file are the same:
>>>> constant viscosity, CFL number=1, same solver tolerance, same spherical
>>>> shell geometry,etc..)
>>>> 
>>>> Case1 takes only 13 timesteps to reach the end time and the end RMS/Max
>>>> velocity is 0.00182/0.0055;
>>>> Case2 takes 20 timesteps to reach the end time and the end RMS/Max
>>>> velocity
>>>> is 0.00558/0.0352;
>>>> Case3 takes 141 timesteps to reach the end time and the end RMS/Max
>>>> velocity
>>>> is 0.0479/0.305.
>>>> 
>>>> They all ran on one node of our cluster but the running time is also
>>>> different:
>>>> case1: 999s
>>>> case2: 1.48e03 s
>>>> case3: 3.86e03 s
>>>> 
>>>> The end min/avg/max temperature and heat flux of upper/lower boundary of
>>>> the
>>>> three cases are all the same.
>>>> 
>>>> From ASPECT manual, I noticed that the time step size is inversely
>>>> proportional to the max velocity of the cell so such a huge difference
>>>> of
>>>> tilmestep size must be caused by the difference of the velocity
>>>> magnitude.
>>>> But in the non-dimemsioanl problem, the same Ra should have the same
>>>> magnitude of velocity. Why in ASPECT, the same Ra can have such
>>>> different
>>>> velocity magnitude? I've realized that ASPECT uses the total pressure
>>>> and
>>>> total density in its momentum equation, so larger value of gravity can
>>>> introduce a larger rho_0*g_0 term in the momentum equation. I'm not sure
>>>> whether this will cause the solver matrix of momentum equation going
>>>> crazy.
>>>> Did anyone else run into this similar confusion? Wolfgang, Timo, Rene,
>>>> Juliane, any insight on this?
>>>> 
>>>> Best,
>>>> Shangxin
>>>> 
>>>> _______________________________________________
>>>> Aspect-devel mailing list
>>>> Aspect-devel at geodynamics.org
>>>> 
>>>> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.geodynamics.org_cgi-2Dbin_mailman_listinfo_aspect-2Ddevel&d=DwIGaQ&c=Ngd-ta5yRYsqeUsEDgxhcqsYYY1Xs5ogLxWPA_2Wlc4&r=R5lvg9JC99XvuTgScgbY_QFS80R7PEA2q0EPwDy7VQw&m=UlGvhK8rWt4pq8DXeR9K1ZYfwj0hBZtV9zNTnj1zJS0&s=uxCinag781DkUL2eMpjkgQodtU35tW1_1Aww6Ngcpww&e=
>> 
>> 
> 
> 
> 
> -- 
> Timo Heister
> http://www.math.clemson.edu/~heister/ <http://www.math.clemson.edu/~heister/>
> _______________________________________________
> 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>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.geodynamics.org/pipermail/aspect-devel/attachments/20170420/432db131/attachment-0001.html>