WRF with Noah-MP: Anomalously Cold T2 Surface Temperature

[jbellino]

Model Code
WRF 4.5.2
Noah-MP v4.5

Platform
Linux x86_64 INTEL (ftn/icc): Cray XC (dmpar)

Model Summary
4km SE USA
Restart run starting from Oct. 1974 through Dec. 2020
Forced with ERA5 from the Copernicus CDS via python API
Geogrid run with GEOGRID.TBL.ARW.noahmp
Notable modules:

• Thompson MP
• CAM radiation physics
• Noah-MP
  • Miguez-Macho & Fan runoff scheme
• BEM urban canopy model (sf_urban_physics = 3)
• FDDA on

Summary of Problem
Anomalously cold computed daily-maximum air temperature at ground level (T2)

Troubleshooting Actions
All troubleshooting was conducted for the simulation period Feb. 11-19, 1984 when there was no anomaly at the beginning of the run and a small one develops after a few days. Feb. 11 coincides with the end of one restart run and the beginning of another allowing us to test whether the model state restart files were tainted. The only action that gave acceptable results was using the Noah LSM instead of Noah-MP.

• run without warm-start from restart files
• Noah-MP: use free drainage and topmodel runoff schemes
• Noah-MP: adjust soiltstep
• Noah-MP: use opt_snf = 4 (wrf precip partitioning)
• Noah-MP: use opt_frz = 2 (Koren's iteration)
• Noah-MP: use opt_stc = 2 (full-implicit soil temp time scheme)
• Try updated code WRF 4.6.0 (solves bug in "GRDFLX" term: Bug fix for Noah-MP snow, vegetation and urban (#1929) · wrf-model/WRF@ce1069f)
• Use Noah LSM

Discussion
Hi all,

I have a 4km model that spans the SE USA that I'm running for the period 1974-2020. In post-processing output I've computed daily maximum surface temperature from T2 (adjusting -5 hours from GMT to EST) and have found significant masses of anomalously cold air at ground level that extends across large swaths of the continental part of my domain and which can last for months on end. One such example begins with the settling of a cold front around the 3rd week of January, 1983 (fig. 1). Left panel shows Tmax from Daymet, middle panel shows calculated Tmax from WRF simulated output with red dots corresponding to GHCN station locations, and the right panel shows the GHCN station-level input observations for the day as published by Daymet (Daymet: Station-Level Inputs and Cross-Validation for North America, Version 4 R1, https://doi.org/10.3334/ORNLDAAC/2132). Model performance is decent enough.

Figure 1. Plots showing Daymet Tmax (left), computed Tmax from WRF simulated output (center), computed Tmax from WRF simulated output versus station-level inputs from GHCN for January 20, 1983.

If we fast-forward to February 16, 1983 (fig. 2) we can see there is a massive amount of cold air (< 45 degrees F) across the northern half of the domain which has not cleared out since the front settled approximately 3 weeks prior. The maps of Tmax are visibly different and model performance is poor. Note the change in scale for temperature.


Figure 2. Plots showing Daymet Tmax (left), computed Tmax from WRF simulated output (center), computed Tmax from WRF simulated output versus station-level inputs from GHCN for February 16, 1983.

Moving ahead in time to April 21, 1983 there are still large pockets of cold air, but we can see clearly that some of it is concentrated along the river valleys in the coastal plain of Georgia and the Carolinas (fig. 3).


Figure 3. Plots showing Daymet Tmax (left), computed Tmax from WRF simulated output (center), computed Tmax from WRF simulated output versus station-level inputs from GHCN for April 21, 1983.

To cap off the demonstration of the extreme nature of the anomalously cold air here is a plot for August 15, 1983 (fig. 4) during which time there shouldn't be daily-max temperatures this low over such a large area of the southeastern US. It's notable that peninsular Florida is relatively unaffected in all of these plots.


Figure 4. Plots showing Daymet Tmax (left), computed Tmax from WRF simulated output (center), computed Tmax from WRF simulated output versus station-level inputs from GHCN for August 15, 1983.

This particular event lasted through December when another front moves through and "resets" everything (fig. 5). Again, note the change in scale for temperature.


Figure 5. Plots showing Daymet Tmax (left), computed Tmax from WRF simulated output (center), computed Tmax from WRF simulated output versus station-level inputs from GHCN for December 25, 1983.

 

[jbellino]

To follow up here with results of some further testing, I tried modifying the &FDDA settings to (1) no longer nudge u,v winds on the full column (i.e. set if_zfac_uv = 1 only nudged above layer 10) and (2) turn off nudging altogether. Both of these modifications had only a small effect on the size of the cold spot that developed and neither prevented it.

 

[Ming Chen]

We are aware of the biases in long-term WRF simulations. Unfortunately we don't have a utility to 'correct' the biases. Note that WRF is originally designed for short-term weather forecast and its climate application is not well evaluated. Please refer to the literature for possible solutions.

 

[jbellino]

Hi Ming Chen, thank you so much for the reply. I understand your comment about biases in long-term simulations; however, I believe that this problem is more than a bias in that (1) these cold air masses linger for months at a time and then suddenly clear out when another front pushes through and (2) I've run the troubleshooting simulations as 10-day forecasts to determine whether the long-term restart run was at issue and found the same cold air masses developing. Unless I can make headway on this in the next couple of weeks, I will likely revert to using the Noah LSM since test results with that module looked ok.

 

[Ming Chen]

Thank you for the update and detailed information.
I will talk to our NoahMP expert and get back to you ....

 

[cenlinhe]

This T2 cold bias in daily max temperature in WRF/Noah-MP is a known issue for the western US, which is partially related to overestimated snow cover and hence surface albedo. In your case, do you also have snow in your study domain during your testing period? If so, you may want to adjust the snow cover parameters (MFSNO and SCFFAC) in MPTABLE.TBL based on the vegetation type in your study domain: noahmp/parameters/MPTABLE.TBL at 848f54ad3d28c4303151fe5ad83724e232694422 · NCAR/noahmp
You can tune these two parameters: increase the values of MFSNO and SCFFAC will reduce snow cover and hence reduce T2 cold bias if your cold bias is occurring together with snow cover in the region.

 

[haiqingsong]

This T2 cold bias in daily max temperature in WRF/Noah-MP is a known issue for the western US, which is partially related to overestimated snow cover and hence surface albedo. In your case, do you also have snow in your study domain during your testing period? If so, you may want to adjust the snow cover parameters (MFSNO and SCFFAC) in MPTABLE.TBL based on the vegetation type in your study domain: noahmp/parameters/MPTABLE.TBL at 848f54ad3d28c4303151fe5ad83724e232694422 · NCAR/noahmp
You can tune these two parameters: increase the values of MFSNO and SCFFAC will reduce snow cover and hence reduce T2 cold bias if your cold bias is occurring together with snow cover in the region.

Could this be happening in other snow-covered areas outside the United States?

 

[meteoadriatic]

Hello,

In NOAH-MP, the cold bias over the areas covered with snow could be improved by changing how the snow thermal conductivity is calculated. You have several options, around line 2565:

! thermal conductivity of snow

  DO IZ = ISNOW+1, 0
     TKSNO(IZ) = 3.2217E-6*BDSNOI(IZ)**2.0           ! Stieglitz(yen,1965)
!    TKSNO(IZ) = 2E-2+2.5E-6*BDSNOI(IZ)*BDSNOI(IZ)   ! Anderson, 1976
!    TKSNO(IZ) = 0.35                                ! constant
!    TKSNO(IZ) = 2.576E-6*BDSNOI(IZ)**2. + 0.074    ! Verseghy (1991)
!    TKSNO(IZ) = 2.22*(BDSNOI(IZ)/1000.)**1.88      ! Douvill(Yen, 1981)
  ENDDO

There, try other models instead of first one. I personally tried constant value (0.35) which, although physically less realistic, seems to me to be much warmer.

I hope this helps.

 

[jbellino]

Hello,

In NOAH-MP, the cold bias over the areas covered with snow could be improved by changing how the snow thermal conductivity is calculated. You have several options, around line 2565:

! thermal conductivity of snow

  DO IZ = ISNOW+1, 0
     TKSNO(IZ) = 3.2217E-6*BDSNOI(IZ)**2.0           ! Stieglitz(yen,1965)
!    TKSNO(IZ) = 2E-2+2.5E-6*BDSNOI(IZ)*BDSNOI(IZ)   ! Anderson, 1976
!    TKSNO(IZ) = 0.35                                ! constant
!    TKSNO(IZ) = 2.576E-6*BDSNOI(IZ)**2. + 0.074    ! Verseghy (1991)
!    TKSNO(IZ) = 2.22*(BDSNOI(IZ)/1000.)**1.88      ! Douvill(Yen, 1981)
  ENDDO

There, try other models instead of first one. I personally tried constant value (0.35) which, although physically less realistic, seems to me to be much warmer.

I hope this helps.

Thanks for this tip! I compiled a new wrf executable after changing that line to use the constant value as you recommended, though it didn't help in this particular instance. I'll keep this in mind going forward!

 

[jbellino]

This T2 cold bias in daily max temperature in WRF/Noah-MP is a known issue for the western US, which is partially related to overestimated snow cover and hence surface albedo. In your case, do you also have snow in your study domain during your testing period? If so, you may want to adjust the snow cover parameters (MFSNO and SCFFAC) in MPTABLE.TBL based on the vegetation type in your study domain: noahmp/parameters/MPTABLE.TBL at 848f54ad3d28c4303151fe5ad83724e232694422 · NCAR/noahmp
You can tune these two parameters: increase the values of MFSNO and SCFFAC will reduce snow cover and hence reduce T2 cold bias if your cold bias is occurring together with snow cover in the region.

Hi Cenlin, thank you so much for the reply. I checked snow cover to see if it was a common factor among multiple of these cold air mass anomalies and, with the limited amount of raw files remaining on the system, I found that it did not seem to be a prerequisite for the condition to develop. I also ran 3 tests in which I increased values for both MFSNO and SCFFAC and I did not see any improvement in the response. I also tried running the tests with fresh input files made with real.exe and the updated MPTABLE.TBL files just to be sure the new values were propagated into the model, but again with no improvement. The last set of values I tried were constants of 4.50 for MFSNO and 0.08 for SCFFAC.

 

[meteoadriatic]

Did you check how much, if any, snow height you have in your cold pool areas? That would narrow down the search for the issue.

 

[jbellino]

Did you check how much, if any, snow height you have in your cold pool areas? That would narrow down the search for the issue.

Unfortunately, our tape backup system is undergoing maintenance and I'm unable to retrieve older raw files containing the SNOWH variable that correspond to the demonstrated case above for Jan.-Dec. 1983. I do, however, have some more recent output for Nov. 2000 which shows a similar, though smaller, cold anomaly (area circled in red) and about 6-12 inches of snow over much, but not all of it. There are also 2 or 3 pixels near Cape Hatteras with more than 4 feet of snow, but this is likely an edge effect near the northeastern boundary. I don't have any snow-depth data on hand to compare to, but I would venture to guess the model is producing too much snow in these areas, particularly for this time of year.

 

[meteoadriatic]

From these plots it seems that cold pool is not caused by snow, at least not directly. Did you had a chance to test different radiation scheme? CAM is a bit unusual choice.

 

[jbellino]

From these plots it seems that cold pool is not caused by snow, at least not directly. Did you had a chance to test different radiation scheme? CAM is a bit unusual choice.

We chose to use CAM a while back when we were running a nested model at 1-km and made a decision to use it because it ran faster than RRTMG. I've seen some slight improvement when running with RRTMG, but the cold air masses are still there just smaller.