Tomorrow.io Cuts AI Weather Forecast Error With Satellite-Fed Starting Data
Tomorrow.io says its Sky 1 assimilation system, fed by 15 microwave sounders, cut Aurora 1.5 forecast error and delivered a roughly one-day lead-time gain over raw GFS initialization.

Why it matters
- Tomorrow.io announced Sky 1 on November 17, updating initial conditions hourly from 15 microwave sounders.
- Adding satellite observations cut initial-state water-vapor error by roughly 5–8% across 28 cases measured against ERA5.
- Sky 1 initialization gave Aurora 1.5 an approximately 24.9-hour lead-time gain over raw GFS initialization against ERA5.
- The sounders carry more than 12 channels spanning 91–204 GHz over a roughly 2,200-kilometer swath.
- 112 Aurora 1.5 forecasts showed lower near-surface temperature error through five days when observations were included.
Tomorrow.io reported on November 17 that better estimates of the current atmosphere improved forecasts from Microsoft's Aurora 1.5 and Google DeepMind's WeatherNext 2 without changing either model. The company's Sky 1 machine-learning data-assimilation system updates initial conditions hourly using observations from its constellation of 15 microwave sounders, which add temperature and moisture information through clouds and over sparsely observed regions.
The stakes sit upstream of model architecture. Most of the current race in weather AI concerns prediction engines — Microsoft's Aurora, Google's WeatherNext — but Tomorrow.io's results argue that observation systems and the software preparing a model's inputs deserve equal attention. Fast inference addresses how quickly a model calculates what comes next; better initialization addresses how well it knows what is happening now. A powerful model can still carry an inaccurate starting estimate forward through its predictions.
What did the satellite comparison find?
Tomorrow.io reports that adding its observations reduced initial-state water-vapor error by approximately 5–8% across 28 cases, measured against ERA5, the ECMWF reconstruction of past weather that combines models and observations into a consistent global dataset.
In the forecast evaluation, 112 Aurora 1.5 forecasts showed lower average near-surface temperature error at every evaluated lead time through five days when the observations were included. The company describes the gains as modest and strongest in the first two days.
The published methods use paired starting states. Sky 1 prepares one with the constellation's observations and another with those observations withheld. The same forecasting model then runs from each, which isolates the contribution of the added data. The station comparison used approximately 17,700 weather stations. WeatherNext 2 was also tested with the paired initial states.
What does the "extra day" result actually mean?
The roughly one-day lead-time gain comes from a separate comparison: Aurora initialized with Sky 1 versus Aurora initialized directly with Global Forecast System (GFS) data. Tomorrow.io reports that the raw-GFS run reached the error level of the five-day Sky 1 forecast approximately 24.9 hours earlier when scored against ERA5. Against weather stations, the equivalent gain was approximately 10 hours.
The company explicitly attributes that result to Sky 1 overall, not to the satellite constellation alone. Part of the improvement comes from giving Aurora an initial state resembling the type it encountered during training.
That distinction matters. The day-equivalent result measures performance at a particular temperature-error threshold and reference. It does not establish that every forecast variable, region or extreme event becomes predictable a full day earlier.
What hardware produces the observations?
Tomorrow.io's microwave sounder documentation describes passive radiometers that supply atmospheric temperature and water-vapor profiles. The company lists more than 12 channels spanning 91–204 GHz, with cross-track scanning over a roughly 2,200-kilometer swath.
These instruments feed an atmospheric analysis; they do not directly output a finished forecast. The value of an observation depends on its calibration, coverage, timing and how the assimilation system incorporates it. That connects the space hardware to the input pipeline used by AI forecasting models.
How do Aurora and WeatherNext fit in?
Microsoft describes Aurora 1.5 as an extension of its Earth-system foundation model, adding hourly temporal resolution, more forecast variables and probabilistic ensemble forecasting. An ensemble represents multiple possible future outcomes and helps users assess uncertainty.
WeatherNext 2 uses a Functional Generative Network architecture for global ensemble prediction. Google's documentation lists a 0.25-degree grid, 64 ensemble members and a 15-day forecast horizon. Tomorrow.io's evaluation covers a five-day horizon.
Google has since introduced WeatherNext 3, which incorporates live satellite observations and produces hourly forecasts. The Tomorrow.io results discussed here concern WeatherNext 2 only; they do not establish an improvement to Google's newer model.
Why does assimilation matter as much as architecture?
A weather model needs an initial state: a three-dimensional estimate of conditions such as temperature, humidity and wind. Observations arrive from different places, altitudes and times. Data assimilation combines those measurements with an existing estimate of the atmosphere to produce a coherent starting point.
ECMWF's explanation of ERA5 describes this combination of models and observations. ERA5 reconstructs past weather into a consistent global dataset, which makes it useful as a reference for forecast evaluation. It remains a reconstruction of the atmosphere, with its own modeling and observation inputs — a caveat that applies to Tomorrow.io's benchmarking as well.
What are the limits of the results?
These are company-reported retrospective evaluations. They provide a useful test of initialization and observation quality, and the release says data is available on request. The announcement describes work evaluating hurricane tracks, atmospheric-river landfalls and severe-weather outbreaks, without establishing the same temperature lead-time gain for those events.
The broader lesson is that weather AI advances through an entire forecasting pipeline: observations, assimilation, prediction and evaluation. Tomorrow.io's tests show how examining the inputs can reveal gains that comparisons of forecasting architectures alone would miss. For users, the meaningful outcome will be reliable improvements in the variables, places and decision windows that matter to their operations — and Sky 1 now demonstrates one path to those gains before a forecast model ever starts running.
Source: unite.ai
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
754 articles
Related articles
- Google DeepMind's WeatherNext 3 Delivers Hourly Forecasts at 5-Kilometer Resolution
- Google DeepMind's WeatherNext 2 Forecasts 8x Faster at Hourly Resolution
- Microsoft Says Its New ML Tool Makes 2-to-6-Week Forecasts Reliable
- Google opens WeatherNext, gaining an extra day to warn of cyclones
- Google opens the datasets behind its WeatherNext 3 forecasts