Theory 13/Forecast error growth/Machine-learned weather models
Error Grows Where the Flow Is Strained
The claim
Forecast error growth rate is set by the ratio of strain rate to error saturation, so errors amplify fastest in deformation zones rather than in the most unstable regions overall.
01Mechanism
Strain stretches error structures until they reach the scale where nonlinear saturation takes over; the ratio sets the growth.
02Prediction
Error growth should peak in strain zones.
03Refutation
Error growth uncorrelated with strain would remove the argument.
04What makes it checkable
Learned forecast models can be run in large ensembles cheaply, giving the error-growth statistics that operational ensembles could not afford at this density.
05Why it was not testable before
Ensemble sizes were limited by compute; cheap emulators allow thousands of members.