Weather model hindcasts stay skillful past 30 days, a study finds, but not yet in real time

The result challenges the long-held view that rapid small-scale error growth limits deterministic forecasts to about two weeks. The authors read it as evidence that the atmosphere may be predictable at large scales for longer than assumed, with predictability perhaps governed mostly by large-scale dynamics.

The work is a hindcast study, not an operational forecast. The optimization needs knowledge of the future trajectory, so the authors say that finding such initial conditions in real time remains an open question and that no operational gains are yet shown.

The team used the small version of GraphCast, which runs on a 1.0° grid and has 36.7 million parameters. They used gradient descent through the model to adjust the two input atmospheric states, widening the optimization window gradually from 2 days in 3-day steps. Single-precision runs covered 14 days; double-precision runs reached 32 days.

Error growth in the optimized forecasts followed a doubling time of 5.8 days from day 2.5 to day 14. After the optimization window ended, error returned toward the control and merged with it near 30 days in single precision and around day 37 in double precision. The authors say this rate should not be read as the atmosphere's intrinsic error doubling time.

To rule out artificial skill from smoothing, the authors measured information error, noise error and forecast activity. Forecast activity stayed close to ERA5 for most variables, which they say shows the improvement comes from better agreement with the verifying field. They also discuss whether the method acts like an adversarial attack on the network and argue it does not, since it minimizes loss and yields physically interpretable large-scale patterns.

The authors say future work could test higher-resolution and coupled atmosphere–ocean models, add regularization, and try the optimized conditions in physics-based models.