search

jbusecke / cmip6_omz

MIT License
2 stars 0 forks source link

Outstanding tasks to rerun density transformation #18

Open jbusecke opened 3 years ago

jbusecke commented 3 years ago

For now you can just use a model that has all output on 'gn' ( I think CanESM5 should work) to test the full pipeline.

sditkovsky commented 3 years ago

I've been going through the pipeline with CanESM5 (r1i1p2f1). Most of it has gone smoothly, but I'm having an issue in the last step; the full volume and omz volume aren't conserved across the vertical transformation to density space. There's about a 60% reduction in total volume between the depth-space and sigma-space oceans. Although areacello and dz_t are conserved individually, their product is not.

The transformation_wrapper code is all the same as for previous projects. And I'm not sure how the changes we have made to the preprocessing steps here would affect this. Did you have this issue before @jbusecke ?

jbusecke commented 3 years ago

That is troubling! Did you make sure to mask the depth values with a mask of where the data is actually present (based on sigma_0 or o2, or even both.

What could be happening here, is that the 'full volume' counts all cells (because it is constructed from a 1D depth, and has no clue about the actual depth of the ocean) and then the transformation will automatically only register dz values where you have density and o2 data. What is the total ocean volume. I did build in some masking steps in my original notebook.

Another issue I encountered was a similar issue with partial bottom cells (but that wont cause a change of 60%). Do you remove the bottom most cells?

jbusecke commented 3 years ago

We could schedule a little debug session tomorrow if you like calendly.

sditkovsky commented 3 years ago

Yes, I'm following the align_missing (checks nan mask consistency between o2 and thetao) and the remove_bottom_values steps from your original notebook. Should removing bottom values deal with the discrepancies between dz and the bottom bathymetry?

The z-space full volume is ~1.28e18 m^3, and the volume in sigma-space is ~4.5e17 m^3

jbusecke commented 3 years ago

~1.28e18 m^3 seems realistic to me. So something is not right in the transformed data.

Should removing bottom values deal with the discrepancies between dz and the bottom bathymetry?

There could be issues with partial bottom cells, but I am unsure they would cause this magnitude of difference.

Can you take a single timestep, integrate the total depth (sum of dz_t) for both depth and density coordinates and plot the difference as a map? I am curious if the differences are close to topography/coasts.

sditkovsky commented 3 years ago

Edit to my earlier comment, areacello isn't conserved globally (and shouldn't be), but I think the issue might be there. It looks like the area is 0 for all of the densest sigma class (I'm using the old pacific-calibrated sigma values for now). It could just be that CanESM5 has a bias against dense water, but looking into this...

Will look at the figure you suggested as well

jbusecke commented 3 years ago

I dont quite understand. Are you transforming the area? I dont think that is necessary. The horizontal area is always fixed and the only thing you need to transform is dz_t.

sditkovsky commented 3 years ago

The area is being transformed because it is cast as a 4-d array before the vertical transformation step (I need to track down exactly where that's happening). And somehow there ends up being no volume/area in the densest bin, even though there are water masses in that density class.

I can try just keeping area as a 2-d array through the pipeline

sditkovsky commented 3 years ago
Screen Shot 2021-07-01 at 5 35 42 PM

Here's the difference in the depth-integrated dz_t for two datasets (z-space minus sigma-space). So it looks like the transformed data tends to be deeper in the interior and shallower on the coasts, but only by 1mm.

jbusecke commented 3 years ago

I just had another idea: Can you add a very low and very high value as the first and last density bin? If you are not counting some of the bins it will not conserve the total depth.

What are the units in that plot? Meter?

jbusecke commented 3 years ago

I think cleaning and progressing to all members of CanESM5 is a good next step. Once we got most of the checkmarks done, we can go into production!

jbusecke commented 3 years ago

Just added match_and_remove_trend and interpolate_grid_labels to cmip6_pp. I will release v0.5.0 today which will include these.

It might still take until tomorrow to get everything through at conda-forge, so maybe just install from source for now.