Different results between adaptive time step and constant dt

[smileMchen]

For the same case running with constant time step and adaptive time step, the results can be quite different. The longer the integration period is, the larger the differences would be. This is a problem that exists at least since WRFV3.8. A test case using WRFV4.2 shows that after 48-hour integration, T2 difference could be within (-4, 4) and rain can be within the range f (-9mm, 9mm). Below is an example that shows the difference in RAINNC after 48 hours of integration.

[davegill]

@smileMchen Ming, This is not unexpected. While lateral boundary conditions will constrain the two simulations from devolving into entirely separate solutions, the small perturbations caused by the time step differences will amplify with the very non-linear moist physics effects.

Notice that the values that you are showing have a distinctive wave pattern (+ - + - + - +, etc). That likely points to not really having a bias, but more of a temporal / spatial offset between the solutions.

What I have seen people do is to run two identical simulations, one with and the other without perturbed ICs. Then look at the resulting differences. You can re-run each of these existing two cases with perturbed ICs with namelist options:

&domains
perturb_input = .true.

Then you will have 4 total cases (two from unperturbed ICs, and two from perturbed ICs). It would be interesting to see those various six pairs of difference fields: C(4,2).

[smileMchen]

@davegill Dave, Thanks for the detailed explanation of the issue. I understand that this is an inherent problem in limited-area modeling. Ming

[davegill]

@smileMchen Ming, It would be nice to see the diffs of the same plotted field RAINNC. If there are objectively larger diffs between the "used constant dt" vs "used adaptive dt" simulations, that is good info for the users to have.