Bug in gravity wave frontogensis code

[jgfouca]

Around line 171 of components/eam/src/physics/cam/gw/gw_front.F90

  do l = 0, ngwv
     where (launch_wave)
        tau(:,l,kbot) = fav(l)
        tau(:,-l,kbot) = fav(l)

Should be

  do l = 0, ngwv
     where (launch_wave)
        tau(:,l,kbot) = fav(l)
        tau(:,-l,kbot) = fav(-l)

[jgfouca]

@rljacob requested that an issue be created.

This will be fixed by #7817

[rljacob]

@jjbenedict or @whannah1 is it possible to say what this bug does physically?

[whannah1]

@jjbenedict or @whannah1 is it possible to say what this bug does physically?

I can't give an "intuitive" explanation about how the change in reynold's stress should affect the model behavior, but I think it can help to think about how this affects the code. This really boils down to the fav variable that is used to set tau.

  if (pgwv > 0) then
     do l = -pgwv,pgwv
        ! Lower bound of bin.
        cmn = cref(l) - 0.5_r8*dc
        cmx = cref(l) + 0.5_r8*dc
        ! Loop over integration intervals in bin.
        fav(l) = 0.5_r8 * dca * (exp(-(cmn/c0)**2) + exp(-(cmx/c0)**2))
...

I think this last line is explained here as this equation:

but it's difficult to verify because there's not a clear 1-to-1 relation between how the variables are used in the code and how they are expressed in that old documentation. I also looked around for papers about this, but couldn't find them. Maybe Yaga Richter or Jack Chen or someone else at NCAR could lead us to the original paper.

Anyway, one critical aspect that relates to the bug is that the reference phase speed array cref has both negative and positive values depending on the sign of the index l

  ! Set phase speeds
  cref = (/ (dc * l, l = -pgwv, pgwv) /)

And the bug effectively makes the code assume that phase speeds are always positive...

However, I've been attempting to demonstrate this more clearly with output from the model and I'm really struggling to find any sort of systematic consequence of fixing this bug.

I did a pair of runs with and without the fix and added a bunch of special outputs for every time step. The plot below shows two of these outputs that represent the zonal wind tendency for a specific range of either positive or negative phase speeds. These quantities have been conditionally averaged based on the instantaneous zonal wind.

My expectation was that we would see a notable difference in the tendencies from negative phase speeds... but that's not exactly the case.

red  = control (no bug fix)
blue = w/ bug fix

The right panel mostly makes sense to me, because we expect negative tendencies (i.e. drag) where a positive phase speed matches the positive zonal wind - which is roughly the condition that causes waves to break. But I don't understand why there is a drag on the negative zonal wind side... I also don't understand why there is a positive tendency in the left panel for positive zonal wind. Most importantly, I don't understand what there is not a big difference between the cases!

I'm gonna try a few more things that might help us explain the effect of the bug fix.

[whannah1]

I found a difference in my tests that might be more directly relevant to this bug.

The variable tau is described to be the Reynold's stress, and there are actually history ouput variables for this in the GW scheme. The variables are divided into each cardinal direction, so in the plot below I've binned the values by the zonal wind, and we can see a notable difference at the larger negative wind values (i.e. westward).

Something I should have mentioned above is that to create these conditionally sampled plots I'm considering every time step over 5-days, but only levels 50-100 mb.

I still don't get how to interpret this result, but I think the difference is directly related to the bug.

[whannah1]

After further consideration I think it's likely that this whole thing is not actually a bug. Furthermore, I think the differences I've found throughout my various tests are just the result of round-off level changes accumulating into seemingly significant changes in model behavior. I'm still not 100% convinced either way, but I've seen enough to convince me that this is not worth our time to investigate for now.