PPE results - Precipitation and LHFLX

[kjetilaas]

Following @adagj example, I am starting a new discussion based on the PPE results, focusing on precipitation and more generally the hydrological cycle.

[kjetilaas]

Broadly we have too much precipitation in NorESM3, and this got worse with b06:

The PPE spans a broad range of global mean precip values, but these follow to some degree the TREFHT, which is too low:

[adagj]

@Ovewh I use the raw data and no current code to do the analysis, and I have sent several emails to everyone with paths to the data and the resources we have, so I don't think there are any a "ppe analysis team" secrets. If you miss files, you can ask @maritsandstad to produce them, or calculate them your self. But no pressure from our side. We are doing our best.

[hermanfugl]

@Ovewh If FLNS isn't among @maritsandstad's global means, you can find global files for three different periods/no. of members here: /scratch/herfugl/diagnostics/ppe_diag/preprocessed_data/ens_var_TS_files

[hermanfugl]

Broadly we have too much precipitation in NorESM3, and this got worse with b06

The PPE spans a broad range of global mean precip values, but these follow to some degree the TREFHT, which is too low

@kjetilaas Is the question whether we can reduce precipitation without reducing temperatures further?

It seems that the cold bias is mainly over land/ice, while the precipitation bias is more global:

NB: 1°C subtracted from the ERA data.

I plotted some regression lines for TREFHT and PRECT over land and ocean, hoping it's possible to pick out tuning choices that adjust the land and ocean (+ Amazon) separately, to some extent:

Immediately, increasing dcs seems like it could lower PRECT globally while having minimal (or positive) impacts on TREFHT over land. Lowering c0_ocn is another option but we're already towards the lower end of what is reasonable(?).
PRECT also seems sensitive to tiedke_add but increasing it seems it could worsen the Amazon bias further.

I should have added the 5 extended members too so we could compare to the plots posted above, but I'm out of time for now.

[maritsandstad]

I can produce FLNS timeseries for everything if that's useful, is it @Ovewh ?

[maritsandstad]

FYI, there is now FLNS timeseries output for all ensemble members @Ovewh

[hermanfugl]

@MichaelSchulzMETNO @benmsanderson the regional precipitation used for the regression plots above can be downloaded as csv here and are saved as netcdf in /nird/datalake/NS9560K/www/herfugl/nc_data -- sorry for the delay, I had to double check a bug.
The PPE parameters can be found in the same directories.

The masks are not perfect, with the oceanmask using a snapshot of the ice edge and the Amazon including the Andes, so keep that in mind:

[MichaelSchulzMETNO]

Are these now annual means of the last year? Or 5 year means?

[MichaelSchulzMETNO]

@hermanfugl sorry, now I see, there are mutliple years in one file, all good

[hermanfugl]

Yes, that's right. Annual means for the 10 years 386--395 in the 21-member ensemble and the 5 years 381--385 for the 76-member ensemble.

[MichaelSchulzMETNO]

What is the mean observed precipitation in the same Amazon region? @hermanfugl

[hermanfugl]

ERA-Interim with the same mask gives 5.87 mm/day.

Global: 2.88, Land: 2.19, Ocean: 3.31, Amazon: 5.87

[benmsanderson]

Here's my quick analysis of the amazon PPE parameter dependencies from the PPE thread.

TLDR: @kjetilaas , maybe try this:
c0_ocn = 0.009
c8 = 4.6
dcs = 0.0006
zmconv_tiedke_add = 0.7

#263 (comment)
#263 (comment)

Constraining by RESTOM, global PRECT and salinity trend gives some scope for tuning amazon precip- but we do cut off the high end:

The parameter responses for global and amazon precip are similar, apart from C0 - where the effect is reversed. Hence reducing C0 should increase Amazon precip, but decrease global precip on average

the actual regression prediction for the amazon is a little noisier than the global case, but still enough to illustrate predictability.

Finally, here's an attempt to maximise amazon precip within the constraints of global precip and RESTOM. The below plot shows the initial global parameter constraint(green) and the top 10th percentile range of amazon precip.

And here's the mean parameter values in that subset of amazon maximising runs:

Parameter Mean Value

micro_mg_dcs 0.000578
zmconv_c0_ocn 0.008922
zmconv_tiedke_add 0.711612
clubb_C8 4.602575
CE 0.065169
EGC 2.025004
EGIDFQ 1.064148
d_max 27.357656
fates_mort_hf_sm_threshold 0.974049
fates_turb_z0mr 1.002037
phi_i_mushy 0.605640

Here's another pespective - we have this trade-off of c0_ocn and c8 which allows us maintain RESTOM=0. within that space, if we choose combinations with low c0ocn and low c8, that tends to increase Amazon precip. Similarly, dcs and tiedke_add should be on the lower end.

mapping back to our original ensemble - run #69 is the best of our current extended simulations, but that actually has middle-of the road values for co_ocn and c8 and there are higher variants in the 15 year set. I think it's possible the mean parameters above might be able to push it up further than anything in the current ensemble.

I think it is possible that we don't have a member which simulataneously maxes out all the degrees of freedom for amazon precip.

[kjetilaas]

Thanks for this, @benmsanderson! I think these values broadly makes sense based on the PPE. But they are actually quite close to the current values in #277, and in fact takes a bit in the wrong direction compared to that:

This is mainly because we already have a low c0_ocn value of 0.0075. I suggest reducing this further, since this is the only parameter that decreases global precip and increases it in the Amazon. But is it a stretch to take this down to 0.005? @oyvindseland @matsbn

This would take us closer to a balanced wrt ocean T, but take precip in the right direction:

[oyvindseland]

I think reducing c0_ocn to 0.005 is fine, but I do not think we should go much further down. It will likely result in more sea-ice however so the ocean parameters need to counter that. c8 can likely also be reduced. It may increase climate sensitivity though. The increase in c0_ocn from CAM5 and onwards compared to earlier versions was done in order to mimic the "iris" effect in the warm Pacific (Warming of SST may decrease high cloud cover so a negative cloud feedback)

[oyvindseland]

Also although I would ideally like to keep the dcs value in the middle of the road I do not think we can go further down than 0.0006. Even that vale will likely give a too low LWCF. tiedke_add may potential counter that.

[Johannesfjeldsaa]

I was just thinking about the usability of the ensemble and so here is a visualization against some observations over AMaZon, Pacific Ocean, Atlantic Ocean and globally. All are from the end of the initial wave (385) uncertainty region for GPCP is +- two stdev computed using annual means. Let me know if you want observations added (but please point to where I can find them).

[oyvindseland]

So basically there is no optimal solution for GPCP global and Amazon. It is likely hard to beat ERA in the first place so is it possible to use ERA for creating optimal solution space, e.g. by using 2 std.deviation for ERA annual means?

[Johannesfjeldsaa]

If you can point me to ERA observations that are not a climatology I'll include it (I used ERAI from adf which "comes" averaged).

[andrearosendahl]

This is ERA5 precipitation from 1979-2020: /datalake/NS9560K/andrear/ERA5_tp_mon_1979-2020.nc

[Johannesfjeldsaa]

@oyvindseland see updated figure. The trade-off between regions are easier to achieve but the amazon low precip is still an issue.

[oyvindseland]

Interesting. It is sometimes hard to see the red dot, but it looks like Kjetil pointed out above. Default is actually one of the better versions.
I notice also that there is a wet bias for most version over both AO and PO. It looks like both PA and PO correlates quite well with global bias so the too salty Atlantic surface vs too fresh Pacific ? may also be a matter of too low transport?