Tomorrow.io refreshes AI weather models' starting point every hour, giving Aurora 1.5 about a day more near-surface temperature lead
Synopsis
AI weather models can produce global forecasts in minutes, but their starting point still mostly comes from conventional data assimilation that updates every six hours. Tomorrow.io's Sky 1 uses machine-learning data assimilation to fold fresh observations from 15 microwave sounders into the initial state hourly. Scored against ERA5, initial water-vapor error in 28 test cases was about 5–8% lower than with those observations withheld; Aurora 1.5 initialized by Sky 1 showed lower station near-surface temperature error at every lead time through five days across 112 forecasts, with modest gains strongest in the first two days.
Interpretation
An AI weather model's initial state can now be refreshed hourly instead of waiting for a conventional data assimilation cycle that supplies a starting point every six hours. Conventional data assimilation systems typically update the estimate of the current atmosphere every six hours, and AI models must wait for that starting point before they can run. Sky 1 builds the same initial state twice, once with Tomorrow Constellation observations and once with them withheld, and the same unchanged model runs from each, so the observations are the only difference.
Folding the constellation's observations into the initial state first makes the initial water-vapor field more accurate, and then makes the downstream AI forecasts more accurate. Without those observations, the initial state draws only on conventional inputs and lacks added coverage over oceans and other sparsely monitored regions. Across 28 cases scored against ERA5, water-vapor error was 5.0% lower on a fresh GFS forecast and 8.0% lower on one eight hours old, 6.7% lower on average over the three to eight hours Sky 1 runs on; across 112 Aurora 1.5 forecasts scored against about 17,700 stations, average near-surface temperature error was lower at all 20 lead times out to five days.
Sky 1 uses machine-learning data assimilation to merge temperature and moisture observations from 15 microwave sounders into the initial state every hour, and these instruments sense through clouds. The microwave sounders add observations over oceans and other sparsely monitored regions that previously lacked fresh atmospheric information. WeatherNext 2 was run unchanged from the same pairs of initial states, with 28 forecasts covering five fields at 13 pressure levels and 20 lead times, showing a pattern similar to Aurora 1.5.
Perspective
The demonstrated scope is 28 initial-state cases, 112 Aurora 1.5 station forecasts and 28 WeatherNext 2 forecasts, referenced to ERA5 and about 17,700 weather stations; effects on hurricane tracks, atmospheric river landfalls and severe weather outbreaks are listed as being evaluated.
The forecast gains are explicitly described as modest and strongest in the first two days; the roughly one-day gain in near-surface temperature lead time is stated to be Sky 1's own result rather than the constellation's, and part of it comes from Sky 1 giving Aurora 1.5 an initial state closer to what it was trained on.
