Upgrade TimeStepWizard to support model spinup?

[ali-ramadhan]

@glwagner I'm thinking of upgrading the TimeStepWizard by adding a spinup_cfl and spinup_time to it, so it keeps CFL < spinup_cfl while model.clock.time < spinup_time. So it would also need a pointer to the model clock.

Does this make sense to you? Would be nice not to have to use two time-stepping loops.

[glwagner]

Consider:

• The Adams-Bashforth time stepper is not well-suited for adaptive time-stepping, because it requires a initialization step. Thus it behooves scientists to update their time-step only infrequently.

• However, during model spinup / transition to turbulence, a sharp change in flow regime may occur. This sharp change necessitates updating the time-step frequently to avoid blow up (an alternative strategy is to take uniform small time-steps during this time, which is safer but will take longer).

• In addition to the need to update a time-step frequently near transition to turbulence, it is also wise to take conservatively small time-steps

Because of these considerations I am not sure it makes sense to add spinup_cfl, which only addresses the last, minor point. The main feature that is needed is a change in the frequency at which the time-step is updated (we currently do this manually by writing two loops). However, a change in update frequency is only really needed because Adams-Bashforth is ill-suited for adaptive time stepping (otherwise it'd be fine to update frequently even in the main loop). So it's a hack-on-a-hack. What do you think?

[ali-ramadhan]

These are good points. I'll close the issue. I hope you don't mind if I use your comments in the documentation under the adaptive time-stepping section.

I ended up finding other ways to transition to turbulence in a stable manner using TimeStepWizard as is, e.g. either increasing the CFL with time

cfl(t) = min(0.005*t, 0.1)
wizard.cfl = cfl(model.clock.time)

or changing the number of time steps between adaptive time steps with time.

[glwagner]

Yes, I’m happy to include general comments. We shouldn’t write too much on the topic. I think our advice will change when we have time stepping methods that are more appropriate for adaptive time stepping.

[glwagner]

Ultimately I think it’d be nice to build adaptive time stepping into the time step function itself, once we have time stepping methods that are well suited for it.