An AI weather model reads the lower stratosphere: wiping eddies above 150 hPa shifts northern Z500 by 55 m at 120 hours
Synopsis
Whether machine-learning weather models such as GraphCast actually use lower-stratospheric and moisture information had lacked controlled input interventions. An audit over 72 ten-day ERA5 forecasts from 2016 to 2021 replaced the atmosphere above 150 hPa with its zonal mean, shifting Northern Hemisphere Z500 by 55.2 m at 120 hours, 1.49 times the response to random-phase noise with identical spectra; global humidity scaling moved 30–70°N Z500 eddy variance monotonically, down to 0.661 of control under complete drying.
Interpretation
A machine-learning weather model's dependence on the lower stratosphere rests on organized eddy structure there, not on perturbation amplitude alone. Earlier input-sensitivity work mostly computed level-wise gradients or ran idealized dynamical tests, without comparing structure destruction against same-spectrum random noise under one norm. Across 72 initial dates, replacing the atmosphere above 150 hPa with its zonal mean shifted Northern Hemisphere Z500 by 55.2 m at 120 hours, 1.49 times the random-phase control response with a 95% interval of 1.40 to 1.59 and Holm-corrected p below 0.001.
Per unit normalized input, a mid-tropospheric perturbation still acts more strongly than the stratospheric knockout, but the gap narrows with lead time. The mid-tropospheric eddy perturbation and the stratospheric knockout had not been placed in a direct ratio under the same norm. The response ratio of the 400–600 hPa random-phase eddy perturbation to the 150 hPa knockout is 4.58 at 24 hours, 2.61 at 72 hours and 1.84 at 120 hours, falling to 1.09 at 240 hours as both approach error saturation.
Global humidity scaling produces a monotonic, sign-reversing dose response in northern mid-latitude eddy variance, unlike a norm-matched temperature-noise control. Earlier moisture interventions concentrated on tropical-cyclone seeds or localized perturbations, without a global multi-dose response curve. At 120 hours the 30–70°N Z500 eddy-variance ratio is 0.661 under complete drying, 0.833 at humidity factor 0.5 and 1.093 at 1.25, while the norm-matched temperature-noise control collapses that variance to 0.215 at 120 hours.
The model recovers from moisture deficits and surpluses asymmetrically, with surpluses removed mainly through precipitation. Recovery times for drying and moistening had not been measured separately in the same model. The total-column-water deficit e-folds in 126 hours at complete drying and at factor 0.25, while surpluses e-fold in 42 hours at 1.25 and 30 hours at 1.5, with about one-third of the 1.5 surplus persisting at 240 hours.
Perspective
The conclusions rest on the public GraphCast-small checkpoint, which has 1° resolution, 13 pressure levels and a 50 hPa model top, was trained on ERA5 1979–2015, and is run for 40 autoregressive 6-hour steps to 240 hours. Evaluation uses 72 initial dates at 00 UTC on the 1st or 15th of each month from 2016 to 2021, with one control and 20 perturbed forecasts per date, and concerns how the model uses lower-stratosphere and tropopause information.
With a 50 hPa model top and a ten-day horizon, the audit does not address the downward influence of sudden stratospheric warmings. Humidity factors above 1 are out-of-distribution inputs, with about 40% of tropospheric mass supersaturated at 1.5; the temperature-noise control is oversized at 8.5 and 4.3 times the natural eddy-temperature amplitude, which the author calls conservative for those comparisons. The drying timescale is defined differently from existing water-vapor lifetime estimates and cannot be equated with them directly.