Should long WRF simulations be interrupted?

[cangyou]

Hi everyone.
I am using WRF to downscale GCM data and simulate summer temperature changes in a city in August 2045-2050. When I try to run WRF continuously for a full month simulation, I get some odd data. However, if I interrupt and restart the simulation (rerun real.exe), the results improve.

My question is, for this type of research, do I need to run one continuous WRF simulation for the full month? Or can I interrupt and restart the simulation periodically? If interruptions are okay, what is a reasonable interval between restarts?

Any advice would be appreciated! I want to make sure I am using best practices for long climate simulations with WRF. Thank you for your help.

 

[Ming Chen]

Since this is a 5-yr simuation, I would suggest that you implement the restart capability of WRF. Note that a "restart" run and a 'continuous' run ( i.e., starting from the initial time and running to the end of your study period) should give you identical results.
If your continuous 1-month simulation yielded different results to that of a 'restart' run, it indicated something went wrong. Please clarify what 'odd' results you obtained from the continuous run.

 

[cangyou]

Since this is a 5-yr simuation, I would suggest that you implement the restart capability of WRF. Note that a "restart" run and a 'continuous' run ( i.e., starting from the initial time and running to the end of your study period) should give you identical results.
If your continuous 1-month simulation yielded different results to that of a 'restart' run, it indicated something went wrong. Please clarify what 'odd' results you obtained from the continuous run.

Thank you very much! In fact, it is not a continuous 5-yr simulation. We only simulate every July in 5 years. And we don't use the 'restart' option. We edit the start and end time in the namelist file, and run real.exe again. Thus the results are different from those of the continuous simulation.
As we focus on the 2-m temperature, we find: there will be rapid cooling and warming on some day, or there will be continuous fluctuations in lower temperatures. Is this due to the numerical oscillation? Thank you again.

 

[Ming Chen]

No, I don't think this is a numerical integration issue. It is more likely that some physics processes lead to the results.

 

[cangyou]

No, I don't think this is a numerical integration issue. It is more likely that some physics processes lead to the results.

Dear Dr. Chen:
Apologies for the late reply, I have been quite busy recently and did not check the forum.
Thank you very much for your insights! We looked into the odd data periods and agree they were likely due to physical processes during rainfall and convection events.

However, we have encountered an issue when running WRF for a one-month simulation driven by FNL data for a historical period. We found that interrupting the run and reinitializing with a 1-day spin-up could improve the simulation, compared to letting it run continuously where errors increase after 10-15 days. Is this behavior normal? Can we run the full month using this restart approach with reinitialization when we want to valuate the model? If so, how would you suggest dividing the simulation into segments for the restarts - perhaps 7 days each?

Thank you again for your guidance! Your expertise is so appreciated.

 

[Ming Chen]

A continuous 1-month run and a run over the same period but with restarting applied during the period should yield identical results. Restarting is not 'reinitializing', instead it is a way to continue the run.

Suppose you run WRF over 1-month period, but you initialize the model every 3-days, then you will have ten 3-day simulations. This is not a 'continuous' run, even if the results cover the entire 1 month.

 

[cangyou]

A continuous 1-month run and a run over the same period but with restarting applied during the period should yield identical results. Restarting is not 'reinitializing', instead it is a way to continue the run.

Suppose you run WRF over 1-month period, but you initialize the model every 3-days, then you will have ten 3-day simulations. This is not a 'continuous' run, even if the results cover the entire 1 month.

Thank you very much. The point is that we found the series of 'reinitializing' simulations (for instance, every 7-days) yielded much better performance than a 'continuous' run (or restarting simulations), when we compared the results with observations. We don't know if it is a normal behavior.... And could we use the 'reinitializing' simulations as validations...

 

[Ming Chen]

This is a normal model behavior, --- model bias will accumulate with longer integration.