Sky 1 gives Aurora a one-day forecast-error gain, Tomorrow.io says
Tomorrow.io says Sky 1, fed by 15 microwave sounders, gave Aurora 1.5 a roughly 24.9-hour lead-time advantage over raw GFS initialization on ERA5 scoring.
By Amara Osei
4 min read

What's News
- November 17: Sky 1 gave Aurora 1.5 a roughly 24.9-hour ERA5 lead-time advantage over raw GFS initialization.
- The equivalent advantage was approximately 10 hours when the forecasts were scored against weather stations.
- Adding Tomorrow.io observations reduced initial-state water-vapor error by approximately 5–8% across 28 cases.
- 112 Aurora 1.5 forecasts had lower average near-surface temperature error through five days.
- Sky 1 updates hourly using 15 microwave sounders with more than 12 channels spanning 91–204 GHz.
By Amara Osei
Tomorrow.io said on November 17 that Sky 1 gave Microsoft’s Aurora 1.5 a roughly 24.9-hour lead-time advantage over raw Global Forecast System initialization on ERA5 temperature-error scoring. Against weather stations, the equivalent advantage was approximately 10 hours, the company said.
The result came from Sky 1 overall, not from Tomorrow.io’s 15 microwave sounders alone. The company said part of the gain came from giving Aurora an initial state resembling the type it encountered during training.
Tomorrow.io’s release says better initial atmospheric data can improve forecasts “without altering the model itself.” Its study covered Microsoft’s Aurora 1.5 and Google DeepMind’s WeatherNext 2 without changing either forecasting model. The company argues that observation systems and software preparing model inputs deserve attention alongside model architecture.
How should the “extra day” result be read?
Tomorrow.io ran a separate comparison between Aurora initialized with Sky 1 and Aurora initialized directly with Global Forecast System (GFS) data. Scored against ERA5, the raw-GFS run reached the error level of the five-day Sky 1 forecast approximately 24.9 hours earlier. Against weather stations, the equivalent gain was approximately 10 hours.
The day-equivalent result measures performance against 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.
Why do weather forecasts depend on starting data?
ECMWF says a weather model needs a three-dimensional estimate of temperature, humidity, wind and other conditions. Its explanation of ERA5 describes data assimilation as combining observations from different locations, altitudes and times with an existing estimate to create a coherent starting state.
ECMWF says ERA5 reconstructs past weather into a consistent global dataset, making it useful as a reference for forecast evaluation. ERA5 remains a reconstruction of the atmosphere, with its own modeling and observation inputs.
Tomorrow.io distinguishes fast inference from better initialization. Inference addresses how quickly a model calculates what comes next; initialization addresses how accurately it represents the present. An inaccurate starting estimate can persist through a powerful model’s predictions.
What did the satellite observations change?
Tomorrow.io reported that adding its observations reduced initial-state water-vapor error by approximately 5–8% across 28 cases, measured against ERA5. The company also reported lower average near-surface temperature error in 112 Aurora 1.5 forecasts at every evaluated lead time through five days when the observations were included.
The company described the forecast gains as modest and strongest during the first two days. The published methods paired two starting states: one prepared with Tomorrow.io’s constellation observations and another with those observations withheld.
The same forecasting model ran from each state, allowing the evaluation to isolate the added data. The station comparison used approximately 17,700 weather stations, and WeatherNext 2 was also tested using the paired initial states.
What feeds Sky 1?
Sky 1 is a machine-learning data-assimilation system that updates initial conditions hourly using observations from Tomorrow.io’s constellation of 15 microwave sounders. The instruments add temperature and moisture information, including through clouds and over sparsely observed regions.
Tomorrow.io describes the sounders as passive radiometers that provide 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.
The instruments supply measurements for an atmospheric analysis; they do not directly output a finished forecast. Tomorrow.io says their value depends on calibration, coverage, timing and how the assimilation system incorporates them.
How do Aurora and WeatherNext fit into the results?
Microsoft describes Aurora 1.5 as an extension of its Earth-system foundation model, with hourly temporal resolution, more forecast variables and probabilistic ensemble forecasting. An ensemble represents multiple possible future outcomes, helping users assess uncertainty.
Google’s documentation says WeatherNext 2 uses a Functional Generative Network architecture for global ensemble prediction. It lists a 0.25-degree grid, 64 ensemble members and a 15-day forecast horizon. Tomorrow.io’s evaluation covers five days.
Google has since introduced WeatherNext 3, which incorporates live satellite observations and produces hourly forecasts. The reported evaluation concerns WeatherNext 2 and does not establish an improvement to Google’s newer model.
What does the evidence establish?
The results are company-reported retrospective evaluations, and Tomorrow.io says the data is available on request. The release describes additional work evaluating hurricane tracks, atmospheric-river landfalls and severe-weather outbreaks, without establishing the same temperature lead-time gain for those events.
Tomorrow.io said, “The broader lesson is that weather AI advances through an entire forecasting pipeline: observations, assimilation, prediction and evaluation.” The meaningful test is whether those gains hold across forecast variables, places and decision windows that matter to users’ operations, rather than a universal extra day for every weather event.
Source: unite.ai
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,336 articles