A weather model learns to correct itself hourly from raw observations

A small machine learning model learns weather forecasting and data blending straight from 18 kinds of instrument. Near the ground it beats the main US ensemble for up to 12 hours, but it trails operational systems higher up.

6 min readIIBenchmarkedPreprint, not peer reviewed
Save
A still life photograph on warm paper showing fine threads connecting scattered small objects to a central wooden disc pinned in blue.AI illustration
Illustration: AI generated

What the study shows

Every weather forecast starts from a best guess of the current atmosphere. Building that guess is called data assimilation: a short forecast is blended with new measurements, each weighted by how much it can be trusted. For global models this is expensive and usually happens every six hours.

The authors, all at the company Zeus AI, built WeatherDreamer to learn that blending step and the forecast together. It is a world model, meaning a model that keeps an internal picture of a system, predicts how it changes, and corrects the picture when new data arrives. It learns from 18 observing systems, from geostationary satellites to weather balloons and about 11,000 surface stations. It never uses reanalysis, the blended historical weather records most AI forecasters train on.

The authors report that at 534 stations hidden from the model, its forecasts of temperature, dew point and wind beat NOAA's global ensemble (GEFS) at every hour up to 12 hours. An ensemble is a set of slightly different forecasts that shows the range of outcomes. Against the European ensemble (IFS ENS) the result is mixed: WeatherDreamer led for temperature only for about the first 8 hours and for dew point only for the first 2.

Its own sense of uncertainty was more realistic. The spread of its forecasts matched the real errors far better than either operational ensemble, roughly twice as closely on the standard ratio used. Sea-level pressure was a clear weakness.

The model also estimated rainfall, sunlight at the ground and cloud-top height without receiving them as inputs. Over the first 6 hours its rain forecasts were the most skilful and the most reliable of the three ensembles, but its rain was too light, about two thirds of observed amounts. Checked against a new satellite it had never seen, its cloud heights were somewhat more accurate than NOAA's operational product.

Compared with NOAA's own data assimilation system on identical observations, its short-range errors were within about 10% at the surface. Higher up they were much worse: up to two and a half times larger for balloon temperatures and more than three times for upper winds. One global hourly update took about 1.5 minutes on one GPU, the type of chip used to run AI models.

Our reading

In our reading, the interesting part is not the headline win over GEFS. It is that a 50-million-parameter model learned the whole predict-and-correct cycle without being told how noisy each instrument is. The authors show its implied trust in some satellite channels lands close to known instrument noise. That is a meaningful sign the model learned something physical, not just pattern-matched.

The forecast wins need careful reading. The authors themselves say much of the advantage comes from the spread of the ensemble, not from a more accurate central forecast. Part of that likely reflects the model predicting directly at station points, while the operational ensembles describe averages over grid boxes. Held-out stations also sit among stations the model did see, and the benefit fades beyond about 200 km from them.

The practical case is clearest near the surface and in the first few hours. A solar farm operator, for example, could get a fresh irradiance and cloud estimate every hour rather than waiting for a six-hourly global update. On readiness, we rate this level II, benchmarked: it is tested on public data against public baselines, but no code or weights are released.

Free membership

Read the rest of this story

Membership is free. You get every article in full and the Issue every Sunday.

We send a sign-in link, no password needed. Already a member? Sign in

Readiness assessment

II
Benchmarked
LevelStatusEvidence
I  Lab onlyMetA working end-to-end system is described in full, with architecture, training details and evaluation on held-out data.
II  BenchmarkedMetBenchmarked on public data against public baselines: GEFS, IFS ENS, GFS, HRRR, NOAA GSI first-guess departures, and GOES cloud-height retrievals. Verification uses ISD, USCRN/SOLRAD, IMERG and EarthCARE. Results are mixed, and this is a preprint.
III  Open sourceNot yetThe paper links no code, weights or model checkpoints. The Open Research section lists only public input data and baseline archives.
IV  In productionNot yetNo documented use by a named organisation. The conflict statement says Zeus AI may use it commercially, but no deployment is reported.

What would move it up
Release the code and trained weights (and the evaluation scripts for the Set A and Set B comparisons) to reach level 3. Level 4 would need documented operational use by a named organisation, for example hourly analyses in a forecasting or energy service, with continuous multi-day cycling and better upper-air and sea-level pressure skill.

The study

Kate Duffy, Elise Penn, Tristan Ballard et al.. WeatherDreamer: Learning a World Model from Multimodal Earth Observations. arXiv, 13 November 2026. Read the paper

Preprint, not peer reviewed · Licence CC BY NC ND 4.0

Notes

  1. Duffy, Penn, Ballard, Posch and Vandal, "WeatherDreamer: Learning a World Model from Multimodal Earth Observations", arXiv preprint, November 2026. Not peer reviewed. Licence CC BY-NC-ND 4.0. Source
  2. No code, weights or checkpoints are linked in the paper. Input observations and baseline forecasts come from public archives listed in the paper's Open Research section.
  3. Exact figures: 10 m wind CRPS averaged over 1 to 12 h was 1.09 m/s for WeatherDreamer, 1.31 for GEFS and 1.15 for IFS ENS. Spread-skill ratio was 0.63 to 0.78, against 0.23 to 0.47 (GEFS) and 0.26 to 0.41 (IFS ENS).
  4. Rain at +3 h (3-h totals): fair CRPS 0.255 mm against 0.310 (IFS ENS) and 0.331 (GEFS). Mean rain rate about two thirds of IMERG. Cloud-top height RMSE 4.4 km against 4.8 km for GOES ACHA; false-alarm ratio 0.09 against 0.19.
  5. Against NOAA GSI first guess (Table 2): held-out surface station departure ratios 1.02 to 1.08; radiosonde temperature 1.79 overall; radiosonde wind 2.62 overall, up to 3.2 to 3.5 at 300 to 200 hPa; ATMS temperature channels up to 4.6 times.
  6. Compute: about 1.5 minutes per global hourly update on one NVIDIA H100; a 12 h, 9-member ensemble forecast takes about 11 minutes. Model has 50 million trainable parameters plus 99 million frozen tokenizer parameters.

Edited and signed by

Stefan PerdalEditor and publisher86 stories edited

Stefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.

How this was made

An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.

More on Vision & multimodal