WeatherDreamer learns data assimilation itself: hourly global analyses in about 1.5 minutes, without reanalysis
Synopsis
WeatherDreamer folds data assimilation and forecasting into one learned cycle trained directly on 18 Earth-observing systems, without reanalysis. It corrects a latent state with new observations every hour; at held-out stations its ensembles beat NOAA's GEFS for temperature, dew point and wind out to 12 hours and are better calibrated, while its assimilation matches GSI near the surface and in moisture channels but shows departures up to about five times larger in temperature-sounding channels. One global hourly analysis takes about 1.5 minutes on a single GPU.
Interpretation
A weather forecast no longer has to start from an initial state supplied by a separate data assimilation system: WeatherDreamer treats the predict-correct cycle itself as something to learn, producing hourly updated global analyses straight from observations. Most machine-learning weather models remain forecast-only engines initialized from separately run data assimilation systems that typically update every six hours and arrive hours late; existing observation-to-observation models build each forecast from a fixed window of observations in a single forward pass and discard the recursive filtering. At 534 held-out stations excluded from the input, its ensembles have lower CRPS than GEFS for 2 m temperature, dew point and 10 m wind at every lead to 12 hours, with spread-skill ratios of 0.63-0.78 against 0.23-0.47 for GEFS and 0.26-0.41 for IFS ENS.
The model keeps a latent state, advances it with learned dynamics each hour, then pulls it back with the observations that arrived during that hour, so background and analysis both live in one hidden space. The cycle follows recurrent state-space models developed for model-based reinforcement learning, with a prior network standing in for the background and a posterior network for the Kalman gain, so observation-error covariances, observation operators and bias corrections are no longer hand-specified. Training and evaluation draw on 18 observing systems, including geostationary infrared radiances, polar-orbiting microwave and infrared sounders, ASCAT ocean winds, GPS radio occultation, surface stations, radiosondes, CloudSat cloud-top height, IMERG precipitation and NSRDB irradiance; CloudSat, IMERG and NSRDB are always withheld as inputs at inference.
Perspective
Evaluation is confined to 50 initializations from 2024-2025 (Set A) and 30 GSI cycles from 2024 (Set B); forecasts are compared against GEFS, IFS ENS and GFS, HRRR is used only for solar irradiance, and station forecasts are compared only with global ensembles; assimilation is compared against the NCEP GSI 3D-Var background. Each inference uses only the preceding 12 hours of observations, and the state is reset to zero at the start of each sequence.
The upper air is the clear shortfall: departure standard deviations for ATMS tropospheric and stratospheric temperature channels are 1.76 and 1.84 times GSI's, reaching 2.7-4.6 times for individual channels, and radiosonde winds reach 3.2-3.5 times at 300-200 hPa; the model draws little from upper-air radiosondes and radio occultation, and its dynamics run on regional patches of roughly 1600 km. Precipitation is dry, at about two-thirds of IMERG's rate, and too smooth, and sea-level pressure is the weakest surface variable.
