Archive of grib2 files containing selected variables from ECMWF IFS ENS. This dataset is currently being backfilled. A cloud optimized version of these data are available at https://dynamical.org/catalog/ecmwf-ifs-ens-forecast-15-day-0-25-degree/ and https://source.coop/dynamical/ecmwf-ifs-ens-forecast-15-day-0-25-degree.
Archives of ensemble weather forecast data from the European Centre for Medium-Range Weather Forecasts (ECMWF).
ECMWF's public open data bucket (ecmwf-forecasts) only provides data from April 2024 onward.
Earlier data is stored in ECMWF's MARS
(Meteorological Archival and Retrieval System), a tape-backed archive that requires careful request
optimization and authenticated API access.
This tool retrieves IFS ENS (Ensemble) forecast data from MARS and stages it as GRIB files in S3, covering the period from 2016-03-08 (IFS cycle 41r2, when ENS resolution increased to 18km) through 2024-04-01 (when S3 open data begins). The staged GRIBs can then be processed by the reformatters into cloud-optimized Icechunk Zarr format, available at dynamical.org.
The files are in GRIB Edition 1 format at 0.25 degree resolution. Each date produces four GRIB files:
Surface variables: sp, 2t, 10u, 10v, 100u, 100v, tp, strd, ssrd, msl, 2d,
ptype. Pressure level variable: z (geopotential) at 500 hPa. Forecast steps: 3-hourly 0-144h
and 6-hourly 150-360h (85 steps total). Ensemble members: 0 (control) + 1-50 (perturbed).
Each date totals approximately 103 GB of GRIB data. The full archive (2,946 dates) is approximately 304 TB.
The directory structure in S3 is:
{s3_prefix}/{YYYY-MM-DD}/{request_type}.grib
For example: dynamical/ecmwf-ifs-grib/ecmwf-ifs-ens/2023-07-01/pf_sfc_a.grib
The staging tool is fully resumable, tracking state in a local SQLite database. It uses ECMWF's
asynchronous MARS API, submitting requests and polling for completion. MARS requests are optimized
to maximize data per tape read: all parameters and forecast steps are requested together (same tape
per date), splitting only by ensemble member count (to stay under MARS's 75 GB retrieval limit) and
by level type (sfc vs pl).
The code is available on GitHub. The compute to retrieve the files is provided by dynamical.org.
Data license: ECMWF data is available under the Creative Commons BY 4.0 license, provided the source is acknowledged.