Using wrfout as wrfinput

Hi,
following up on topic >Difference between wrfinput and initial wrfout

I observed no significant difference between wrfinput and the initial wrfout in WRF4.2.2 for the test case em_quarter_ss.
Yet, I found a case where there is a difference:

When I integrate the test case for 1,5 hours and use the wrfout file there as wrfinput to restart the run, like so: then I get a significantly different P field in the wrfinput and first wrfout file at this time.

Is it not advisable to restart a run from a wrfout file?
I heard that it should not make much difference. Yet I am concerned about the error that is introduced by that, since I get T errors oscillating around 1K (RMSE, lowest level, comparing wrfout_restarted run to a wrfrst_restarted run; see figure below)

Ming Chen said:

If I understand correctly, the case that shows large difference is a restart case ( https://forum.mmm.ucar.edu/phpBB3/viewtopic.php?f=40&t=10031#p20222)

Note that one cannot restart a WRF run based on wrfout. This is because many intermediate variables required for restarting are not included in wrfout. One must use wrfrst to restart a WRF run.

Click to expand...

Thank you for your quick response!
I understand that it is not possible to restart a WRF run based on wrfout (in the sense that it is a bit-reproducible run).
Yet, it is possible to initialize a run with a wrfout file, and I was advised that this is the way to create a wrfinput file after data assimilation (not with WRFDA, but with DART).

So, what is the recommended way to do data-assimilation?
Do you advise against initializing a WRF run after a data-assimilation step by renaming wrfout_d01_ to wrfinput_d01?
Or do you recommend inserting updated states into a wrfrst file instead? (ncks -A -v da_output wrfrst_d01_)

Many thanks!

Hi Lukas,

WRRF can be initialized in two ways:

(1) cold start: initial condition is derived from large scale forcing data like ERRA-I, HGS, etc.
(2) run data assimilation to produce initial condition

We often run cold start for research issues, while people in operational service prefer to run assimilation for real-time forecast.
It all depends on your specific cases.