UChicago AI Weather Model Reforecast Archive

Data documentation

Overview

A multi-model archive of long-range reforecasts from AI weather models, initialized 2000–2025 and rolled out to 50 days — beyond the models' traditional forecast lengths — at daily cadence on a global 0.25° grid.* The archive supports benchmarking of AI weather models, training of downstream ML models, and multi-model blending. It also powers AI-Almanac, UChicago-LAUDE's interactive, human-centered forecast platform. Real-time forecasts initialized from operational ECMWF IFS analyses are planned alongside the reforecasts.

Models

Model Class Grid Members Forecast length
GraphCastdeterministic721 × 1440 (0.25°)50 d
Aurora 1.5deterministic720 × 1440 (0.25°)50 d
AIFS-single v2deterministic721 × 1440 (0.25°)50 d
AIFS-ENS v2ensemble721 × 1440 (0.25°)5150 d
FGNensemble721 × 1440 (0.25°)5150 d
Aurora 1.5 ENSensemble720 × 1440 (0.25°)5150 d
NeuralGCMensemble64 × 128 (≈2.8°)5150 d

Planned updates

The archive is maintained as a living resource. New AI weather models are added as they are released — we expect roughly one to three additions per year, tracking the current pace of model releases.

Each new model is run under the same protocol as those already in the archive: the same 2000–2025 initialization dates, 50-day rollouts, and variable save set. Newly added models are directly comparable to the existing ones without additional reprocessing.

Archive structure & format

Forecasts are stored as one zarr store per model per initialization: forecasts_<model>/init_YYYYMMDDT00.zarr. Dimensions are (time, [number,] prediction_timedelta, lat, lon) for 6-hourly fields and prediction_timedelta_daily for daily aggregates; number is the ensemble-member dimension. All data are float32.

Daily aggregates are given by: means of the four 6-hourly values in the UTC calendar day; precipitation is the daily accumulation; 2 m temperature max/min are daily extremes.

Acknowledgements

The reforecasts were generated on the DSI Cluster at the University of Chicago.

Licensing & attribution

Forecasts were produced with third-party models. The archive does not relicense them; the terms below apply to the weights that generated each store.

Provider Model Code Model weights
ECMWF AIFS-single v2 Anemoi, Apache 2.0 CC BY 4.0
ECMWF AIFS-ENS v2 Anemoi, Apache 2.0 CC BY 4.0
Google DeepMind GraphCast Apache 2.0 CC BY 4.0
Google DeepMind FGN (WeatherNext 2) Apache 2.0 CC BY 4.0
Google Research NeuralGCM Apache 2.0 CC BY-SA 4.0
Microsoft Research Aurora 1.5 MIT MIT
Microsoft Research Aurora 1.5 ENS MIT MIT

ECMWF data

Copyright © European Centre for Medium-Range Weather Forecasts (ECMWF). Source: www.ecmwf.int. Licence: CC BY 4.0. ECMWF does not accept any liability whatsoever for any error or omission in the data, their availability, or for any loss or damage arising from their use. Contains modified Copernicus Climate Change Service information. Neither the European Commission nor ECMWF is responsible for any use of the Copernicus information or data.

These reforecasts are experimental machine-learning output, produced for research use. They are not official products of any model provider or meteorological agency.

* NeuralGCM is the exception, on a ≈2.8° grid.