Auditing how an AI weather model uses lower-stratospheric and moisture information: structure-destroying interventions in GraphCast

Abstract We audit how GraphCast (1°, 13 levels, top at 50 hPa) uses lower-stratospheric and moisture information via input interventions in 72 ten-day ERA5 forecasts (2016–2021, 18 per season). Replacing the atmosphere above 150 hPa with its zonal mean changes Northern Hemisphere Z500 by 55 m at 120 h, 1.49 [1.40, 1.59] times the response to incoherent random-phase noise with identical per-latitude spectra. Zonal-mean and climatological replacements are indistinguishable over all dates. Per unit normalized input, mid-tropospheric perturbations act 4.6 times more strongly than the stratospheric knockout at 24 h (6.6–6.9 under alternative normalization) and 1.8 times at 120 h; the stratospheric share is largest in boreal winter. Global humidity scaling changes 30–70° N Z500 eddy variance monotonically, down to 0.66 of control for complete drying, unlike a conservative norm-matched temperature-noise control. Moisture deficits recover with an e-folding time of about 126 h; surpluses e-fold within 30 h, but one-third persists.