Reproducing model air sea fluxes with aerobulk

[jbusecke]

Given the progress we made over at aerobulk-python I wanted to try the algo out on the CM2.6 data just for a quick test.

I was able to use the xarray wrapper to successfully compute air-sea fluxes for a month and a subset of the domain.

When I compute the airseaflux for daily data using the 'ecmwf' algo (presumably the one used in CM2.6; personal comms with Steve Griffies), and then averaging all values in time, I can compare the value with the output of the model (which is provided as monthly average).

Unfortunately the difference is pretty darn big:

These show a percentage difference compared to the model output.

Some notes:

• The difference in the Arctic/Antarctic are very large! This is likely due to sea-ice? We can probably ignore that for now.
• But what could explain the still large differences (10-30%)?
  • Could this be due to the fact that we do not have the skin correction activated? I tried to dig through the [CM2.5 code]https://www.gfdl.noaa.gov/cm2-5-and-flor-quickstart/) but wasnt able to find the option to activate/deactivate the skin correction.
  • This could be compounded with the fact that we only have daily averages available to reconstruct the fluxes, and as such dont resolve the diurnal cycle?
  • What I did here to get the atmospheric fields onto the ocean is a simple linear interpolation to the ocean grid. But since the errors show quite a fine structure, I would assume the issue is primarily with the SST

Overall I think this needs to be addressed before we can make any confident statement about the scale separation. Let me know if I missed something here.

[rabernat]

Other possibilities:

• double check z_t, and z_u,
• niters parameter

[rabernat]

Here is the order of operations I would follow to debug this

1. Verify that the results of aerobulk-python are numerically identical to those of aerobulk fortran. This means excavating the aerobulk fortran test data, running their test scripts, etc. This step is important because it allows us to credibly raise any issues we find upstream in aerobulk. Otherwise, the aerobulk developers can always just say that our package is broken somehow.
2. Run an MITgcm simulation with a prescribed atmospheric state (e.g. global_with_exf) for a single timestep. Verify that we can reproduce the air-sea flux offline with aerobulk-python.
3. Find a full set of climate model outputs (atmosphere and ocean) for which we have hourly outputs from both ocean and atmosphere and understand exactly which air-sea interaction scheme (+ parameters) was used. Verify that we can reproduce the air-sea flux offline with aerobulk-python.
4. Only at this point could we confidently investigate whether diurnal cycle, monthly averaging, skin temperature, etc. etc. are responsible for this discrepancy.

[jbusecke]

Ok so 1. Is actually implemented in test/test_fortran.py of aerobulk_python now!

Quick update: I reran the above analysis with the skin correction on (just because it was relatively easy to do), and the results are definitely better!

But I think its not the full story, and following @rabernat steps above seems like the right avenue!

[paigem]

Thanks for retesting this with the skin correction @jbusecke! You're right - definitely better, but not good enough. I'll put this on the agenda for our Monday meeting to discuss our next steps based on @rabernat's suggestion above.

[jbusecke]

Some updates:

I originally did get the zt and zu heights wrong. CM2.6 provides the atmospheric velocities at 10m:

And temperature and humidity at provided at 2 m:

I have adjusted the defaults on the xarray wrappers in https://github.com/xgcm/aerobulk-python/pull/33

I am afraid this did not really improve the fit between the native model fluxes and the recomputed ones.

The search goes on.