Different results between adaptive time step and constant dt

[wrf_alessandro]

Dear All,

I am running WRF3.8.1 for some hindcast simulations and I am getting different results between a simulation with fixed dt and a simulation using adaptive time step.
I do not expect to get identical bit-to-bit results, but differences are quite large for both surface air temperature and accumulated rainfall (see attached figures), also considering that I force the model to write output at the same time:

adjust_output_times = .true.
step_to_output_time = .true.,

I wonder if anyone here faced the same issue or anybody can provide me some hints to understand what is going wrong?
Thanks into advance for help,
Alessandro

 

[Ming Chen]

Alessandro,
Would you please run the same case with and without adaptive time step option using WRFV4.2? Please let me know whether large differences still exist between the two options. I ask this because we may not debug problems in older version of WRF. But it is a problem in the latest version of WRF, we definitely will look into it. Thanks.

 

[wrf_alessandro]

Ming,
I ran WRF4.2 with a slightly different set-up and I still have the same issue (please look at the attached figures); at this point I guess this is not related to WRF version or physical parameterizations. Any other suggestions?
Many thanks,
Alessandro

 

[Ming Chen]

Would you please upload your wrfinput, wrfbdy and namelsit.input files for me to take a look? I will repeat this case first. Thanks.

 

[wrf_alessandro]

I have just uploaded on nextcloud a file called FORCING.tgz containing namelist.input, wrfinput, wrfbdy for a 5-day run.
Thanks,
A.

 

[Ming Chen]

Hi,
I am trying to download the file you uploaded. Unfortunately the file size is too large that it froze my computer.
Would you please just attach your nameless.input and namelist.wps, I will manage to produce wrfinput and wrfbdy.
Thanks.

 

[wrf_alessandro]

Ok, no worries, please find attached the namelist.wps and namelist.input I used for the run. By the way, I am using ERA5 to force WRF.
Thanks,
Alessandro

 

[Ming Chen]

Alessandro,
I am writing to let you know that I can repeat the problems you reported. The longer the integration, the more severe the problem could be.
Unfortunately I haven't figured out what is wrong. We will look into this issue and post fixes in our future release of WRF. Thank you for reporting this problem.

 

[wrf_alessandro]

Hi Ming,
thanks for cross checking and confirming the issue. So, should I assume that adaptive time step is not working properly, while results from constant dt simulation are reliable?

 

[Ming Chen]

Alessandro,
I went through all the codes related to fixed and adaptive time step. I didn't find anything unreasonable. I talked to our expert about this issue and we believe that the differences are attributed to the small perturbations caused by the time step differences, which later amplify with the very non-linear effects of model physics and dynamics.
By the way, have you ever examined the results and know for sure which option gives you better results? Thanks.

 

[wrf_alessandro]

Ming,
I did not performed a sensitivity analysis to parameterizations, but I used different physical schemes between WRF3.8.1 and WRF4.2 and the outcome is roughly the same. I understand the issue could be attributed to the small perturbations, however a 2C difference for temperature is very large; besides, considering for accumulated fields, in case of long simulations the differences become impressive. Would you suggest to use the adaptive time step anyway?
Thanks,
Alessandro

 

[Ming Chen]

Alessandro,
If possible, please stay with the fixed time step.
I am actually perplexed by the large differences between fixed and adaptive time step, which is not what we expect. Yet except for the non-linear impacts caused by small perturbations, we have no explanation for this large difference. I will pay attention to this issue, and if we indeed find something , I will let you know.