Negative promotion/demotion area due to double float precision

[jiamenglai]

Describe the issue

Hello, I met an error of 'negative or more area than the cohort has is being promoted/demoted' in EDCanopyStructureMod, and I think it's due to double float precision. I am running ELM-FATES with ed_st3 on and with my inventory file. The trigger of this error is in the code lines below (Line 662 - 664):

remainder_area = cohort%c_area - layer_co(ic)%pd_area
copyc%n = cohort%n*remainder_area/cohort%c_area
cohort%n = cohort%n - copyc%n

When I output those values, I found that pd_area is about 4E-16 (positive), but the copyc%n is slightly larger than the original cohort%n, which is wrong mathematically. This gives a negative cohort%n after those code, and led to negative demotion area in the next demotion phase.

Any suggestions to fix this bug? Thanks!

Relevant log output

FATES tag

sci.1.88.0_api.42.0.0

Host land model tag

v3.1.0-alpha-4304-ge909ef9f6f

Machine

perlmutter

Other supported machine name

No response

Additional context

No response

[rgknox]

Thanks for bringing this up @jiamenglai . I'm currently digging into this part of the code as well, after seeing some similarly strange behavior related to zero cohort number densities generated in this area. I'll keep this thread involved if/when I figure anything out. Very helpful, thank you!

[mpaiao]

Thanks for making this issue @jiamenglai. I don't know if it is helpful, but maybe you could paste the extra output you added to figure out the problem in this issue too.

[jiamenglai]

Thanks @mpaiao for the suggestion. I revised the code to output some variables:

remainder_area = cohort%c_area - layer_co(ic)%pd_area
! The copied cohort is the part that remains in-layer
copyc%n = cohort%nremainder_area/cohort%c_area
if (copyc%n > cohort%n) then
write(fates_log(),)'copyc%n > cohort%n'
write(fates_log(),)'cohort%n',cohort%n
write(fates_log(),)'copyc%n',copyc%n
write(fates_log(),)'pd_area',layer_co(ic)%pd_area
write(fates_log(),)'remainder_area',remainder_area
write(fates_log(),*)'remainder_area/cohort%c_area',remainder_area/cohort%c_area
endif
! The original cohort changes layers
cohort%n = cohort%n - copyc%n

The output is as below:

[rgknox]

@jiamenglai I submitted a potential fix here: #1509

If you could test this branch and let me know if it fixes your problem, or if there are ways we can improve it?

[mvdebolskiy]

@rgknox
Do you commonly test with comp_excln_exp<0?
I've encountered even worse (though with NorESM code-base and parameter file, I'll check that paramfile on derecho with whatever ESCOMP/CTSM checks out later):
FPE in:

fates/biogeochem/EDCanopyStructureMod.F90

Line 584 in 56add1b

layer_co(ic_nn)%pd_area = remainder_area*layer_co(ic_nn)%p%c_area/group_area

I assume it's another case of tiny numbers making numerical instabilities.

[jiamenglai]

@jiamenglai I submitted a potential fix here: #1509

If you could test this branch and let me know if it fixes your problem, or if there are ways we can improve it?

This branch works for my case. Thanks!!