# WeatherBench 2 GraphCast Forecasts (ERA5-initialised, 2018 and 2020)

## Facts

- Provider: Google DeepMind / Google Research (WeatherBench 2)
- License: CC-BY-4.0 (https://storage.googleapis.com/weatherbench2/datasets/graphcast/LICENSE)
- Format: Zarr (in sampled files)
- Stored at: Google Cloud Storage, gs://weatherbench2/datasets/graphcast (us)
- Access: public, readable without an account
- DOI: https://doi.org/10.1126/science.adi2336
- Category: Climate & Weather
- Source: https://doi.org/10.1126/science.adi2336
- Page: https://scigantic.com/archive/3e1c3c0e-3831-5939-9429-0df6183327e5
- Keywords: Climate & Weather, weather forecasting, graph learning, benchmark, Google Cloud Storage

## About

Deterministic medium-range global weather forecasts from GraphCast, the graph neural network weather model of Lam et al. (Science, 2023), as published in the WeatherBench 2 benchmark. Forecasts are initialised from ERA5 every 12 hours and run out to 10 days in 6-hour steps (40 lead times, 6 h to 240 h). The 2018 forecasts come from a model trained on data up to 2017 and the 2020 forecasts from a model trained up to 2019, so each year is an out-of-sample test. Measured 2026-11-14: the 2020 store spans initialisation times 2019-11-16 to 2021-01-31 (886 times) and the 2018 store 2017-11-16 to 2019-01-31 (884 times), on the native 0.25 degree 1440x721 grid with 37 pressure levels from 1 to 1000 hPa. Variables include geopotential, temperature, specific humidity, u and v wind, vertical velocity, 2 m temperature, 10 m wind, mean sea level pressure and 6-hour total precipitation. Each year also has regridded 240x121 and 64x32 conservative versions and a _derived store; the full-resolution 2020 store is roughly 33 TB uncompressed. Files are consolidated Zarr; open with xarray.open_zarr and score against ERA5 with the weatherbench2 package. Licensed CC BY 4.0 by Google LLC; cite Lam et al. 2023 and Rasp et al. 2024 (WeatherBench 2).

## Load the data

```python
import gcsfs
fs = gcsfs.GCSFileSystem(token='anon')
fs.ls('weatherbench2/datasets/graphcast/', detail=False)[:10]
```

Lists the files with anonymous access. Runs on any machine, with no cloud account.

## Data structure

File types in sampled files: Zarr

Example file: `datasets/graphcast/2018/date_range_2017-11-16_2019-02-01_12_hours-240x121_equiangular_with_poles_conservative.zarr/.zgroup`

Path pattern: `{prefix}/{seg1}/{YYYY}/{seg3}/{filename}`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/3e1c3c0e-3831-5939-9429-0df6183327e5
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/3e1c3c0e-3831-5939-9429-0df6183327e5
- Over MCP: `get_schema_card` with id `3e1c3c0e-3831-5939-9429-0df6183327e5` at https://api.scigantic.com/mcp

## Related topics

- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [Climate & Weather datasets](https://scigantic.com/topics/climate-and-weather)
- [Weather forecasting datasets](https://scigantic.com/topics/weather-forecasting)
- [Graph learning datasets](https://scigantic.com/topics/graph-learning)
