Open dorchard opened 1 year ago
Do we need to try to match the SAM parameters (https://github.com/yaniyuval/Neural_nework_parameterization/blob/v.1.0.3/sam_code_NN/sam_cases/run_files_x8_5_layers/prm) to the component set and grid config for CAM (sounds reasonable...)
So far I've got reasonably far with this setup:
./cime/scripts/create_newcase --case ~/Documents/iccs/my_cesm_sandbox/dom.test --compset QSPCAMS --res f19_g17 --machine homebrew --run-unsupported
but need input data now..
@dorchard 2.2 is still in development according to NCAR with 2.1.3 being the most recent production release. Advice seems to be to stick with 2.1.3 - we could start on this and then move to 2.2 once we have something that works in stable and hop it isn't too large a change?
The docs for CESM seem to suggest there is 2.1, 2.2, and 'upcoming/beta', however, it's not clear if 2.1 or 2.2 is the latest stable.
The latest tagged release on the main branch of the CESM github is 2.1.3.
From the meeting today we ascertained that we want v2.1
I changed the name now of this to focus on the configuration!
Note that Paul suggested we configure CESM to use the SCAM component (Single Column Atmospheric Model) ~at data point twp06~ based on the observational data (related to this paper: https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018MS001578
As a further update @paogorman suggests the using gateIII testcase for tropical convection in SCAM (see paper in above comment) as it does not involve land (which twp06 does).
We also discussed model resolution to which he commented: "One issue we were concerned about for using SCAM as a testbed for the NN parameterization is the implied horizontal resolution. SCAM takes T42 as the resolution input. My reading of the documentation is that this T42 resolution enters because of the advection scheme which is based on an older dynamical core with that horizontal resolution. Since SCAM is a single column model and it will only use vertical advection, I don’t think it matters for our purposes. There is also the issue of the effective resolution of the large-scale forcing data from observations used by SCAM. For the IOP cases we will consider, the observations were taken on an effective grid of greater than 96km, so in the long run we might want to switch to the x16 parameterization. But I think for now it is fine to stick with the default NN parameterization (x8) which is the best documented in the paper and on github."
See SCAM test case setup here: https://ncar.github.io/CAM/doc/build/html/users_guide/atmospheric-configurations.html#cam-single-column-fscam-compset
We probably want to have a 96km grid config (see https://www2.cesm.ucar.edu/models/cesm2/config/grids.html) maybe C96_C96_mt061 ?
We think that the following includes the right case setup for us this:
% cd cime/scripts
% ./create_newcase --case test_scam_gateiii --compset FSCAM --res C96_C96_mt061 --user-mods-dir ../../components/cam/cime_config/usermods_dirs/scam_gateiii
% cd test_scam_gateiii
% ./case.setup
% ./xmlchange STOP_N=600
% ./case.build
% ./case.submit
Do you agree @paogorman and @judithberner? I am not certain about the grid configuration though.
(for my local setup I also add --machine homebrew --run-unsupported
to the create_newcase
)
Hi Dominic, looks good to me. I normally start with running only a few timesteps, ./xmlchange STOP_N=10, but this model is quite cheap, so it won't matter much/ Let me know how you get on, Judith
On Tue, Apr 18, 2023 at 6:25 AM Dominic Orchard @.***> wrote:
We think that the following includes the right case setup for us this:
% cd cime/scripts % ./create_newcase --case test_scam_gateiii --compset FSCAM --res C96_C96_mt061 --user-mods-dir ../../components/cam/cime_config/usermods_dirs/scam_gateiii % cd test_scam_gateiii % ./case.setup % ./xmlchange STOP_N=600 % ./case.build % ./case.submit
Do you agree @paogorman https://github.com/paogorman and @judithberner https://github.com/judithberner? We are not certain about that grid configuraiton though.
— Reply to this email directly, view it on GitHub https://github.com/m2lines/convection-parameterization-in-CAM/issues/4#issuecomment-1506723221, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADIZOG32SOMIITBRK4BLDPLXB2B3JANCNFSM6AAAAAAV2UM3DI . You are receiving this because you were mentioned.Message ID: @.*** com>
-- Judith Berner https://staff.ucar.edu/users/berner Email: @.*** Scientist III Mesoscale and Microscale Meteorology (MMM) Laboratory and Climate and Global Dynamics (CGD) Laboratory
National Center for Atmospheric Research Boulder, CO 80305, USA
Potentially useful in future so making a note here: It is possible to run SAM using SCAM input files in the latest version should we wish to do a direct comparison.
I am re-opening this issue as the move to Derecho has raised some problems.
On Cheyenne we used CESM version 2.1.0, but this is not building on Derecho (due to python issues I think). I have successfully built v2.1.5 and v2.2, but when I try and build using the SourceMods I updated for v2.1.0 it no longer works, and a diff shows substantial differences in the source.
Before I get too far ahead do you have any recommendations for the version to work with on Derecho going forward @paogorman @judithberner ? Assuming the ZM/deep convection hasn't changed my feeling is to use v2.2.0 as reasonably recent?
Digging a bit deeper it seems the latest science release of CESM is v2.1.5 which we should probably use. As a result I will proceed with updating our modifications for the version of CAM associated with that if that seems reasonable?
As a side note, the version of CAM associated with CESM v2.2.x has a very different interface compared to what we have been modifying, so will require significant changes to the SourceMods files to use.
We know we want to be running CESM in "aquaplanet" mode. Looking at the quick start guide we configure CESM via the
create_newcase
tool. Which then leads to questions:[ ] Which component set definitions should we use (see https://www2.cesm.ucar.edu/models/cesm2/config/compsets.html) for aquaplanet mode. There are 18 component sets when searching here for "aquaplanet"
[ ] Which grid resolution?
[ ] Which input data should we use (can we use something smaller and coarser?)