Why does the training time step is different from the time integration steps? Why does neuralGCM learned physics tendencies recomputed once every 30/60 minutes while "gradually increase the rollout length from 6 hours to 5 days"?

[weatherforecasterwhai]

Hi,
First, in the Training part of the paper, it clearly said### "We gradually increase the rollout length from 6 hours to 5 days".
But, in the ### appendix E time integration,### " learned physics tendencies are only recomputed
once every 60 minutes for our lowest resolution (2.8◦) model and every 30 minutes for all others."

Why does the training time step is different from the time integration steps?

And I tried, if setting
inner_steps = 1 # save model outputs once every 24 hours
timedelta = np.timedelta64(1, 'h') * inner_steps
it also works.
And when testing p_e, for the sam time preriod, the 1 hour step running are the same with 6 hours running.

Thank you.