AIWP Pangu-Weather GFS Reforecast

GribStream Code: aiwppangu

Official source: The AI Weather Prediction archive is a collaboration between CIRA and NOAA Global Systems Laboratory for evaluating data-driven global weather models. This page covers Pangu-Weather forecasts initialized from NOAA GFS analyses.

Current GribStream coverage: aiwppangu begins with the May 27, 2025 00 UTC cycle and grows forward as twice-daily forecasts arrive. The upstream archive contains older cycles beginning in October 2020, but cycles before May 27, 2025 are outside this GribStream dataset's coverage.

Experimental research availability: The registry warns that data may be missing and availability is not guaranteed; the public README says the data are provided as-is and documents earlier publication-timing and directory-structure changes. This evaluation dataset is not an official operational NOAA forecast. Do not make it the sole dependency for availability-sensitive systems: monitor forecasted_at freshness and keep an operational fallback dataset.

Version continuity: Each GribStream code is pinned to the current source-model version. A materially different upstream model version will receive a separate dataset code rather than being silently joined to this history.

Description

Pangu-Weather is a data-driven global weather prediction model. The AIWP project runs it from GFS analyses to create a consistent archive for community evaluation. AIWP deliberately starts its Pangu-Weather history after the model's training period, helping researchers study out-of-sample forecasts rather than examples the model learned from.

The original Pangu-Weather system represents pressure levels as a true third spatial dimension. Its 3D Earth-specific transformer is designed to learn relationships between atmospheric states at different heights, while its hierarchical temporal aggregation combines networks trained for 1, 3, 6, and 24-hour advances to reach medium-range lead times with fewer iterative steps.

In the original Pangu-Weather paper, the model was trained on ERA5 from 1979 through 2017 and tested on 2018. Against reanalysis, the authors reported lower error and higher anomaly correlation than the comparison systems for all 69 evaluated factors, and studied tracks for 88 named tropical cyclones. Those research results used different initial conditions and evaluation periods, so they are context for the model—not a performance guarantee for this current GFS-initialized archive.

The four near-surface fields describe 2 metre temperature, 10 metre wind, and mean sea-level pressure. Five upper-air variables describe geopotential, temperature, horizontal wind, and specific humidity at 1,000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150, 100, and 50 hPa. Together they support analysis of large-scale circulation, thermal structure, moisture, and storm evolution. This product does not include precipitation, clouds, radiation, or surface fluxes.

GribStream exposes the published variable names, levels, values, and units unchanged so results remain traceable to the source. The AIWP team asks users to cite the AIWP archive paper. For operational decisions, compare this research output with authoritative current guidance such as GFS and NOAA's operational AIGFS products.

Detail

• Archive begins: 2025-05-27
• Official AIWP dataset registry: https://registry.opendata.aws/aiwp/
• Key features:
  • Complete 1,440 by 721 global grid: 1,038,240 cells at 0.25 degree latitude-longitude spacing.
  • Twice-daily 00 and 12 UTC cycles with 41 forecast steps from 0 through 240 hours at 6-hour intervals.
  • Nine variables and 69 exact selector tuples: u10, v10, t2, msl, plus z, t, u, v, and q on 13 pressure levels from 1,000 to 50 hPa.
  • Pangu-Weather uses a three-dimensional Earth-specific transformer to model relationships across pressure levels and a hierarchy of 1, 3, 6, and 24-hour prediction networks to limit error accumulation.
  • Variable names, pressure levels, float32 values, and published units remain unchanged; the generated inventory documents every available selector.
• Resolution: Regular global 0.25 degree latitude-longitude grid with 1,440 longitudes by 721 latitudes, ordered from 90N to 90S and covering 0 through 359.75 degrees longitude in the source.
• Temporal coverage: Cycles at 00 and 12 UTC. Each cycle contains forecast hours 0, 6, 12, and every 6 hours thereafter through 240, for a 10-day horizon.
• Primary uses: Controlled AI forecast evaluation, comparison with GFS and other global models, synoptic-pattern analysis, temperature and wind studies, moisture transport, tropical-cyclone research, forecast verification, renewable-energy studies, route planning, and machine-learning benchmarks.

Related datasets

• GFS for NOAA operational numerical guidance and a conventional baseline
• AIGFS Surface for NOAA operational AI forecast fields
• AIGFS Pressure for NOAA operational AI pressure-level fields
• GraphCast GFS for another historical GFS-initialized AI forecast archive

Parameters