Small basal velocity value patch from Adjoin inverse method

Extension of Elmer in computational glaciology
ygong
Posts: 15
Joined: 08 May 2013, 16:51
Antispam: Yes

Small basal velocity value patch from Adjoin inverse method

Hi,

I've currently used adjoint inverse method to do the velocity inversion and got some patches with very low basal velocity (almost 0) in some fast flow area (see the pics). I haven't got the similar thing when I used Robin inverse method. Since there are some 0 values in the observation I wonder if this is just the way how the adjoint method treats the 0 value.

P.S. the value of the cost function looks fine.

Cheers,
Yongmei
Attachments
The velocity,beta and costfunction value
Last edited by ygong on 10 Jan 2014, 17:47, edited 1 time in total.
ygong
Posts: 15
Joined: 08 May 2013, 16:51
Antispam: Yes

Re: Small basal velocity value patch from Adjoin inverse method

There is a pic with better scale that you can see the patch more clearly. Observational surface velocity on the left and modeled basal velocity on the right.
Attachments
velocity.pdf
Velocity with better scaling
gagliar
Posts: 77
Joined: 04 Sep 2009, 16:34
Location: LGGE - Grenoble
Contact:

Re: Small basal velocity value patch from Adjoin inverse method

Hi Yongmei,

No value or a value = 0 is completely different. If you have no value for the surface velocity, it should not be imposed as a zero value. You should impose velocity only where you have measured value.

Hope it helps
Olivier
Martina
Posts: 39
Joined: 17 Apr 2013, 14:06
Antispam: Yes

Re: Small basal velocity value patch from Adjoin inverse method

Hi Yongmei,

I'm not sure if I understand you correctly:
If you have zero velocities in your observed velocities in fast-flowing areas I'd rather suspect that this an problem in your input data than a problem with your run. How else would you explain these slow points in the middle of your fast flowing outlet? If these very slow velocities are really in your data, it's only a good sign if your adjoint model reproduces them! However, I'm a bit puzzled that I cannot see these spots in your figure of the data. Are they really there? Why are they more pronounced in your modeled figures?

I'd say that you should have a closer look at your data and understand these small values. And then either replace them by your no-data-value (if you have one) and impose the velocity as Olivier suggests only where you have input data; or delete them and fill the arising data gaps with interpolation. (If I remember right there should already be scripts around to do such interpolation; most likeliest it could simply be done when reading the data in Elmer with pointwise or whatever Solver you use to read your ascii data!)

Martina
tzwinger
Posts: 97
Joined: 24 Aug 2009, 12:20
Antispam: Yes

Re: Small basal velocity value patch from Adjoin inverse method

Hi,
I am not that sure if there's something wrong. As we discussed today and from your latest picture you can deduce that you are comparing the measured velocities at the surface with the computed at the bedrock. The match at the surface (not shown here) was visually quite OK and also your L-curve was acceptable. IMHO - and may anyone in the Forum please correct me, if make a mistake - there can be sticky spots even in between fast flow regions at the bed and it still can be a good solution, as long as the free surface velocity matches.

Best Regards,

Thomas
ygong
Posts: 15
Joined: 08 May 2013, 16:51
Antispam: Yes

Re: Small basal velocity value patch from Adjoin inverse method

Hi,
Thank you very much!

I think I confused poeple a bit that my 0 value in surface obeserve are not originally in those fast flows but somewhere else I didn't show (very few of them). But the basal velocity produced by Adjoint do have those patches.

I don't quite understand how Adjoint mehtod can numerically produce those 'sticky spot' (however I don't have them when I use Robin inverse method with the same surface velocity data) ?

Cheers,
Yongmei
Martina
Posts: 39
Joined: 17 Apr 2013, 14:06
Antispam: Yes

Re: Small basal velocity value patch from Adjoin inverse method

Hei Yongmei,

if your small surface velocities are elsewhere, I don't see a link between them and your sticky spots.

I think it is perfectly possible that you find different solutions with the Robin method and the Adjoint method - it's different ways of doing the inversion and if you take third one, you'll certainly find a third solution for your basal drag.
I would just make sure that what you see is not an artefact of the choice of your paramter in the regularisation term (in addition to check that everything converges fine - not only the Stokes Solvers, but also m1q)? Did you vary the regularisation parameter?
You could also try another initial condition for beta to exclude some problems in the minimization.

I agree with Thomas that there could be sticky spots - especially when "only" using an inversion and not any physically based model for the sliding.

I see two things to check and think about:
- I would definitely have a close look on the observational data in these regions to exclude that there is any error or strange feature in this data which could be the origine of these sticky spots. (Maybe do the same for your bedrock/surface data/mesh.)
- From what Thomas writes, I understand that the surface velocities are fast over the whole outlet without slower spots.
How can it be that the surface velocities match the observations, i.e. are fast over the whole outlet; but the bedrock has sticky spots? That means that in these regions you have a high surface velocity which apparently is not caused by sliding (otherwise the bedrock velocity also would be high). Can deformation account for this high velocity at the surface or where does it come from? Or - if the high surface velocity should be due to the sliding - how can it be that the bedrock velocity is so small? This difference between bedrock and surface velocity described by Thomas puzzles me much more than the existence of the sticky spots as a such.
Maybe what you see is something like what Denis mentioned a while ago - he had higher velocities at the basis than at the surface in some points. That's also something what I do not really understand. [But I haven't seen your modeled surface velocities, so maybe this is again a misinterpretation.]

Martina
Last edited by Martina on 13 Jan 2014, 18:55, edited 1 time in total.
ygong
Posts: 15
Joined: 08 May 2013, 16:51
Antispam: Yes

Re: Small basal velocity value patch from Adjoin inverse method

Hi Martina,

Thank you very much! I'll have a closer look at my data.And I just put my modeled surface velocity here.

Cheers,
Yongmei
Attachments
surface velocity 95.pdf
surface velocity