Implement option to decenter the application of physics tendencies to achieve higher global accuracy

[KevinViner-NOAA]

Is your feature request related to a problem? Please describe. Application of physics tendencies only at the end of the dynamics represents a first-order accurate coupling. By applying the physics tendencies on the right-hand-side of the small step dynamics, we can achieve near second-order accurate coupling. This has been demonstrated in semi-Lagrangian models like IFS and UM. This may also solve some of the strong variation in solutions we see in RRFS when varying the time step size.

Describe the solution you'd like I want to put in a parameter ("phys_decenter") which applies a linear combination of physics tendencies from the previous time step on the RHS of the dynamics and physics tendencies from the new time step at the end of the dynamics. Using a value of phys_decenter=0.50 represents a second-order coupling scheme.

Describe alternatives you've considered There are likely other ways to increase the order of accuracy of the coupling, but this method allows us to: 1) maintain the physics as a single suite (a la CCPP), 2) is relatively simple to implement, 3) can recover the original configuration by setting phys_decenter=1.0 PDC in UFS.pptx

Additional context Add any other context or screenshots about the feature request here.

[lharris4]

Hi, Kevin. Thanks for putting this together. This could be a very useful addition to the UFS.

Rusty and I have discussed this a bit. We believe it may be a better idea for CCPP to optionally return tendencies (rather than the state) for which the decentering could already have been applied. Then all of the blending is handled within CCPP rather than needing to be concerned about adding new functionalities to FV3, which supports a number of different modeling systems.

Lucas

On Tue, Dec 6, 2022 at 9:44 AM KevinViner-NOAA @.***> wrote:

Is your feature request related to a problem? Please describe. Application of physics tendencies only at the end of the dynamics represents a first-order accurate coupling. By applying the physics tendencies on the right-hand-side of the small step dynamics, we can achieve near second-order accurate coupling. This has been demonstrated in semi-Lagrangian models like IFS and UM. This may also solve some of the strong variation in solutions we see in RRFS when varying the time step size.

Describe the solution you'd like I want to put in a parameter ("phys_decenter") which applies a linear combination of physics tendencies from the previous time step on the RHS of the dynamics and physics tendencies from the new time step at the end of the dynamics. Using a value of phys_decenter=0.50 represents a second-order coupling scheme.

Describe alternatives you've considered There are likely other ways to increase the order of accuracy of the coupling, but this method allows us to: 1) maintain the physics as a single suite (a la CCPP), 2) is relatively simple to implement, 3) can recover the original configuration by setting phys_decenter=1.0 PDC in UFS.pptx https://github.com/NOAA-GFDL/GFDL_atmos_cubed_sphere/files/10167423/PDC.in.UFS.pptx

Additional context Add any other context or screenshots about the feature request here.

— Reply to this email directly, view it on GitHub https://github.com/NOAA-GFDL/GFDL_atmos_cubed_sphere/issues/230, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMUQRVEVX6DA6ELTO2CSF33WL5GIDANCNFSM6AAAAAASVSZOMQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[bensonr]

@KevinViner-NOAA @lharris4 - adding a little to Lucas' comment. Our proposed approach would also make the capability directly available to all modeling systems which utilize CCPP without the need for modifications in their broader infrastructure.

[KevinViner-NOAA]

@lharris4 @bensonr I'm not sure I understand your proposed solution. This method is applying the physics tendencies on the RHS of the acoustic step, I'm not seeing how that's possible through CCPP. The best we can do without changing the FV3 model is "dribbling" at each remapping step; this changes code one level higher and adds physics tendencies at each k_split. It's unfortunate, but I think we're stuck with low order physics-dynamics coupling if we maintain the modularity of physics and dynamics.

[lharris4]

@KevinViner-NOAA the higher order physics implementation is certainly worth trying, and if it is useful it would be a valuable addition to FV3. (I'm not particularly fond of physics-dynamics modularity and indeed one of the goals of the FV3 Integrated Physics project is to break this strict separation.)

Could you give us some more implementation details, like which codes need to be modified in FV3?

Thanks, Lucas

[climbfuji]

@lharris4 @KevinViner-NOAA I haven't been able to follow the conversation and what problems there are w.r.t. to the CCPP physics interface, but if it helps we can have a chat/call anytime soon to go over things.

[ligiabernardet]

I am also interested in staying in the loop in this discussion, especially wrt future needs for development of CCPP or phys-dyn interface for the UFS in general.

[KevinViner-NOAA]

@lharris4 @bensonr @climbfuji @ligiabernardet Here's the code I've been using to test the decentering, please let me know if you have any questions or see anything incorrect: https://github.com/KevinViner-NOAA/GFDL_atmos_cubed_sphere/commit/48aba3c751f44818de8476a90f9b1cfc5bc60b55

[lharris4]

Hi, Kevin. I have a number of concerns about the decentering implementation:

• The tendencies should be an optional argument to fv_dynamics, especially since in the adiabatic na_init initialization step they would necessarily be 0. They also are not supported in other FV3-based models that use a different atmosphere.F90.
• I really don't think it is necessary to bring them all the way down to the acoustic (n_split) timestep; applying the decentered tendencies after vertical remapping (k_split timestep) would make more since, as the levels have been remapped to the Eulerian reference levels on which the physics is executed. This is only if the dribbling is considered necessary, which wasn't part of the original decentering proposal.
• The tendencies for the winds should be remapped to the D-Grid only once, so that we don't need to spend flops to re-do the remapping.
• I don't believe that a separate array for qv needs to be used since water vapor is the first tracer.
• Do beware of modifying the tendencies for tracers, since they need to be applied correctly to ensure mass is conserved and to avoid rounding error loss of mass in single-precision dynamics.

[bensonr]

@KevinViner-NOAA - did this ever get addressed in a PR?