Fix bg_gradient lh bug #4111
Fix bg_gradient lh bug #4111
Conversation
| get_qi_sat(thermo_params, bg_model) * | ||
| TD.latent_heat_sublim(thermo_params, t_sat) | ||
| ) / | ||
| (get_ql_sat(thermo_params, bg_model) + get_qi_sat(thermo_params, bg_model)) |
There was a problem hiding this comment.
add an eps to make sure this is not divided by zero?
There was a problem hiding this comment.
Good call. I thought I would be able to avoid it since we are only in that function if we have some non-zero cloud condensate. But I guess the numerics were not playing well with me - I actually need and sqrt(eps(FT)) in the denominator to make the implicit Rico single column test case work
There was a problem hiding this comment.
sorry, I meant max(..., eps(FT)), but I can change it later. (or does max not work?)
There was a problem hiding this comment.
Did it change the results of any of the CI? I'm not super familiar with all the plots.
Also was wondering if it should converge to a real L_v at q=0 instead of 0, not that it should make a huge difference in the presence of any real cloud. I guess in equilibrium, condensate converges to 0 at the same time as the supersaturation, but that's less true in noneq.
sqrt(eps(Float64)) is only 1.49e-8 which is not unheard of for thin ice clouds, and if we ever ran Float32 simulations is 3e-4 which isn't insignificant (though I don't know if that's intended to be possible)
When I trialed fixing bg I just set to lh to Lv at ql+qi < eps(FT) -- I guess you could also just add +eps(FT)L_v or something similar to the numerator and denominator if you wanted it to be more numerically stable and continuous.
There was a problem hiding this comment.
prognostic edmf dycoms in ci changed. There are probably some other changes too. And we always run Float32. I agree we should do this more carefully. Could you work together with @trontrytel to make this better?
There was a problem hiding this comment.
Oh, I have an idea. We can use TD.liquid_fraction here? That way Lh is always between Lv and Ls.
trontrytel
force-pushed
the
aj/buoyancy_gradients_lh
branch
from
7858c8f to
1a40a8c
Compare
This PR fixes the latent heat bug discovered by @jbphyswx in our buoyancy gradients.
It would be great to rewrite parts of that file to get rid of the getter functions at the top and the bg_model, and to reduce the number of functions we define. But maybe we can do it in the next PRs.