LMD/KPP performance in river plume simulations

[dylanrs617]

Hi All,

I'm doing a model-model comparison of the Texas-Louisiana shelf with ROMS and MPAS-O. Our ROMS model is validated (Hetland 2017 JPO, Kobashi & Hetland 2020 OD) and normally uses k-omega or k-epsilon for vertical mixing with great success. The mean resolution is ~1.5 km. 30 vertical layers are stretched towards the surface (theta_s=5). Only a small amount of explicit harmonic mixing (5.0 m^2/s and 1.0 m^2/s) that is scaled to the grid size is applied to momentum/tracers to destroy grid-scale noise. MPAS-O uses KPP configured as part of the CVMix library (GLS isn't available) and I think part of the reason we see big differences between models is that KPP is struggling, which it is known to in shallow regions with interacting boundary layers. My understanding is that KPP can be sensitive to the time-stepping & advection schemes and the literature has shown significant differences between the ROMS implementation of KPP and CVMix.

I've done a few experiments with the ROMS model (since I can't test this with MPAS-O) but have had little success with KPP. See below for a comparison of the surface normalized vorticity and temperature during the first summer of spinup. The GLS (k-omega) simulation is consistent with our previous work and looks great. Only difference between the two is the vertical mixing scheme. You can see that KPP does not produce a submesoscale eddy field and the Mississippi near field is saturated with grid scale noise. The KPP simulation is also too diffusive, which I confirmed with probability density functions of the vertical viscosity & diffusivity (not shown). I've attached the relevant parts of the log file for the KPP simulation, but the key CPP flags are below.

Code: Select all

 LMD_BKPP                 KPP bottom boundary layer mixing
 LMD_CONVEC               LMD convective mixing due to shear instability
 LMD_MIXING               Large/McWilliams/Doney interior mixing
 LMD_RIMIX                LMD diffusivity due to shear instability
 LMD_SHAPIRO              Shapiro filtering boundary layer depth
 LMD_SKPP                 KPP surface boundary layer mixing

I've done a trial simulation with biharmonic mixing, but that made little difference in the grid scale noise and didn't improve the representation of the eddy field. My next idea is to turn off RI_SPLINES and try smoothing the Richardson number. Has anyone had success using KPP in realistic simulations of river plumes? Any suggestions?

Thanks,
Dylan

[dylanrs617]

Just for an update, I found significant improvement with Richardson number smoothing paired with biharmonic mixing. See attached for a vorticity comparison with the same GLS simulation (that uses harmonic).

Code: Select all

 RI_HORAVG                Smooth Richardson number horizontally
 RI_VERAVG                Smooth Richardson number vertically
 TS_DIF4                  Biharmonic mixing of tracers
 UV_VIS4                  Biharmonic mixing of momentum

and the following mixing coefficients as a first cut with back of the envelope estimates:

Code: Select all

 1.0000E+06  nl_tnu4(01)       NLM Horizontal, biharmonic mixing coefficient
                                 (m4/s) for tracer 01: temp
 1.0000E+06  nl_tnu4(02)       NLM Horizontal, biharmonic mixing coefficient
                                 (m4/s) for tracer 02: salt
 1.0000E+06  nl_tnu4(03)       NLM Horizontal, biharmonic mixing coefficient
                                 (m4/s) for tracer 03: dye_01
 1.0000E+06  nl_tnu4(04)       NLM Horizontal, biharmonic mixing coefficient
                                 (m4/s) for tracer 04: dye_02
 1.0000E+07  nl_visc4          NLM Horizontal, biharmonic mixing coefficient
                                 (m4/s) for momentum.

Next, I will do an ensemble to optimize the amount of horizontal mixing since the fronts in KPP are a little weaker. I haven't looked into the vertical tracer structure yet, but this is big improvement. Hope this helps someone in the future!

Dylan

[jpringle]

Have you looked at this patch to the code, and the ROMS forum pages linked to from it? https://github.com/myroms/roms/pull/51

They are likely relevant to you issues, because they discuss overmixing with GLS.

Jamie

[dylanrs617]

Thanks for sending this. I was unaware of the bug - but I am using COAWST v3.8, which has ROMS 4.1. I will need to see how difficult it is to git cherry pick the commit and deal with merge conflicts.

That said, I don't think GLS is overmixing in this case (recall I paired it with harmonic lateral mixing). My impression is that KPP is still too diffusive, but it is still quite close to GLS. See below for probability density functions of normalized vorticity during summer for GLS and KPP comparison with different values of biharmonic viscosity & diffusivity. The diffusivity is set to be an order of magnitude smaller than viscosity. Since these coefficients are scaled to the grid size - most values are about 10 times smaller than listed in the plot. Even when reducing del4 by an order of magnitude, the tails of the distribution don't change much. The peak is slightly offset, so I will continue tuning experiments, but since the lateral mixing is not affecting the frontal strength much I think this is more related to KPP.

Dylan