WeatherBench 2

WeatherBench 2 is an open benchmark and curated data archive from Google Research for evaluating data-driven global weather models over the medium range (one to fifteen days ahead). It updates the original WeatherBench (Rasp et al., 2020) and bundles three things: an open-source evaluation framework, publicly hosted ground-truth and baseline datasets, and a continuously updated leaderboard that ranks models against ECMWF's IFS.

The data lives in a public Google Cloud bucket (gs://weatherbench2/datasets) in cloud-optimised Zarr, so you can stream slices without downloading everything. It includes a comprehensive copy of ERA5 reanalysis (1959 to 2023, six-hourly) as ground truth, climatologies, IFS analysis, and forecast archives from GraphCast, Pangu-Weather, FuXi, NeuralGCM and others, provided at several resolutions from 0.25 degree down to coarser grids such as 1.5 degree.

For AI work it is best read as an evaluation harness rather than a text corpus. The arrays are numeric grids of temperature, wind, geopotential and precipitation, so you cannot embed them directly; the practical route for RAG is to pair them with the framework documentation, model cards and the accompanying paper, or to turn specific queries into short factual passages. Its real strength is reproducible scoring: standard metrics, fixed ground truth, and a shared leaderboard.

On licensing, the benchmark code is Apache-2.0, while the underlying ERA5 data carries the Copernicus/ECMWF licence, which permits commercial use with attribution and applies no territory exclusion. The caveat is that individual baseline forecast datasets ship their own LICENSE files and some restrict commercial use, so check each before building a product on it. Where Copernicus and NASA Earthdata are broad earth-observation portals, WeatherBench 2 is the narrow, AI-ready evaluation layer for weather forecasting specifically.