change edmfx advection

[trontrytel]

Closes #2106

I wasn't sure if we also want to change the tke advection?

[trontrytel]

@tapios , @dennisYatunin - Could you take a look. Below I'm showing results from the edmfx advection test. The first plot shows results for the min_area equal to 1e-2 and the second to 1e-3.

• The spikes in updraft buoyancy, potential temperature, etc disappear only when min_area is set to 1e-2, which is pretty large.
• There is a secondary spike in the updraft q_tot (below 5km) that gets larger as min_area gets larger. I don't have an explanation for it right now.
• The grid mean q_tot doesn't seem to be equal to the area weighted average of the drafts? - Look for example at the point where updraft area is the largest.

TODO - next step: check hat rho^0*q_t^0 + hat rho^1*q_t^1 vs rho q_t on the gm

TODO - next next step: Add the weight for updraft and environment to output. Plot it and see how it looks in problematic areas.

ClimaAtmos.jl/src/utils/variable_manipulations.jl

Lines 85 to 92 in 5f855e5

	function divide_by_ρa(ρaχ, ρa, ρχ, ρ, turbconv_model)
	weight = sgs_weight_function(ρa / ρ, turbconv_model.a_half)
	# If ρa = 0, we know that ρa / ρ = 0, which means that weight = 0. However,
	# 0 * ρaχ / 0 = NaN, regardless of what ρaχ is, so the linear interpolation
	# will always return NaN when ρa = 0. To avoid this problem, we need to add
	# a special case for ρa = 0.
	return ρa == 0 ? ρχ / ρ : weight * ρaχ / ρa + (1 - weight) * ρχ / ρ
	end

min_area = 1e-2

min_area = 1e-3

min_area = 1e-4

min_area = 1e-4 Float64

[trontrytel]

Here is another clue. I'm plotting the difference between the grid mean ρq_tot and the area weighted contributions from the drafts computed in two ways: Updraft contribution based on the state Y and based on the cache p.

You can see that the cache based version has a big spike where we have the problematic secondary peak in q_tot.

Updraft computed from the state: (Y.c.ρq_tot - Y.c.sgsʲs[1].ρaq_tot - p.ᶜρa⁰ * p.ᶜspecific⁰.q_tot):

Updraft computed from the cache: (Y.c.ρq_tot - Y.c.sgsʲs[1].ρa * p.ᶜspecificʲs[1].q_tot - p.ᶜρa⁰ * p.ᶜspecific⁰.q_tot):

[trontrytel]

Here is how the weight = sgs_weight_function(ρa / ρ, turbconv_model.a_half) looks like for updraft and environment

ClimaAtmos.jl/src/utils/variable_manipulations.jl

Lines 64 to 71 in 0d2ce17

	sgs_weight_function(a, a_half) =
	if a < 0
	zero(a)
	elseif a > 1
	one(a)
	else
	(1 + tanh(2 * atanh(1 - 2 * (1 - a)^(-1 / log2(1 - a_half))))) / 2
	end

[tapios]

Here is another clue. I'm plotting the difference between the grid mean ρq_tot and the area weighted contributions from the drafts computed in two ways: Updraft contribution based on the state Y and based on the cache p.

Looks like state and cache are out of sync. May this get better with @charleskawczynski's refactoring of the precomputed quantities to postcomputed quantities?

[szy21]

Looks like state and cache are out of sync. May this get better with @charleskawczynski's refactoring of the precomputed quantities to postcomputed quantities?

Maybe. It's more likely from that ρaq_tot is not equal to ρa * q_tot from the regularization I think.

[tapios]

Looks like state and cache are out of sync. May this get better with @charleskawczynski's refactoring of the precomputed quantities to postcomputed quantities?

Maybe. It's more likely from that ρaq_tot is not equal to ρa * q_tot from the regularization I think.

Can we run this test with just the limiters on entrainment/detrainment I sketched, without any additional regularization?

[charleskawczynski]

Maybe. It's more likely from that ρaq_tot is not equal to ρa * q_tot from the regularization I think.

Agreed, the cache is currently in sync, the only difference we should see from moving precomputed to post computed quantities is that we should be able to reduce the number of calls by 1 (technically + a fraction due to reducing in callbacks), while keeping the state and cache in sync.

[trontrytel]

Can we run this test with just the limiters on entrainment/detrainment I sketched, without any additional regularization?

This test is only with advection. There is no entrainment/detrainment here

[tapios]

Can we run this test with just the limiters on entrainment/detrainment I sketched, without any additional regularization?

This test is only with advection. There is no entrainment/detrainment here

So entrainment=detrainment=0?

[trontrytel]

So entrainment=detrainment=0?

Yes. The test was meant to look only at advection issues. No other tendencies are applied.

I think that the issue is in the regularization function. I'll try to track it today

[tapios]

So entrainment=detrainment=0?

Yes. The test was meant to look only at advection issues. No other tendencies are applied.

I think that the issue is in the regularization function. I'll try to track it today

Sounds good. Maybe the test is more meaningful with small but nonzero entrainment, detrainment, with the limiter functions on them, to ensure that area fraction remains bounded and updraft properties relax to grid means.

[trontrytel]

Can we merge this change as it is? No matter what we do with the regularization, we still want to switch our advection.

Merging it in would help me with testing for other cases. cc: @tapios @szy21

[szy21]

Sounds good to me.

[trontrytel]

bors r+

[bors]

Build failed (retrying...):

• buildkite/climaatmos-ci

[trontrytel]

bors r+

[bors]

Already running a review

[bors]

Build succeeded!

The publicly hosted instance of bors-ng is deprecated and will go away soon.

If you want to self-host your own instance, instructions are here.
For more help, visit the forum.

If you want to switch to GitHub's built-in merge queue, visit their help page.

• buildkite/climaatmos-ci
• docbuild
• format

[trontrytel]

Below are results with/without using grid mean density Y.c.rho when doing upwind reconstructions for edmfx variables. I'm running the test without any moisture, to highlight that the problem is not with the phase change. The last set of plots is the advection test for our target vertical resolution and using the min_area = 1e3.

• Changing the density in the advection reconstructions doesn't change the results much.
• When switching to the target resolution and increasing the min_area, the problem is less severe.

I'm going to merge those settings as the default for prognostic edmfx and move on to the next test case (GABLS).

Using grid mean density in upwind reconstructions:

Using updraft density in upwind reconstructions:

Target vertical grid: