dClimate/aifs-single-2-earthboi

AIFS Single 2.0 on earthboi

ECMWF's AIFS Single 2.0, the model behind ECMWF's operational AIFS forecasts since 12 May 2026, run by the earthboi platform from its IFS initial conditions and scored against station observations.

This repo holds no weights. earthboi.yaml declares ECMWF's repo, which the platform mounts read-only at /weights, pinned to the commit main pointed to when this wrapper was registered. predict.py does to the platform's initial conditions what ECMWF's run_AIFS_v2.0.ipynb does to open data:

1. reads the T-6h and T0 states: 14 pressure levels of z, t, u, v and q (q from 100 hPa), 13 surface and soil fields including snow depth, the 11 ocean-wave fields, and 4 constants;
2. interpolates them onto AIFS's N320 grid with the matrix earthkit-regrid uses, turns mean wave direction into its cosine and sine, and sets snow depth and soil water to NaN over sea with ECMWF's lsm.grib;
3. runs the forecast with anemoi-inference to +168h;
4. interpolates 2 m temperature back onto the 0.25 degree grid for scoring.

The network is closed while it runs, so the two interpolation matrices ship in regrid/. make_regrid.py fetched them from earthkit-regrid 0.5.1, and regrid/SHA256SUMS pins them.

IFS initial conditions only. Snow depth and the wave fields are not in the ERA5 file, so a run from ERA5 fails with a message saying so.

Runs on the torch-2.7 runtime on a 24 GB GPU; it peaks at about 6 GB.

How close it is to ECMWF's own run

On init 2026-09-14T00Z, run the notebook's way from the same analysis, 2 m temperature came within 0.18–0.24 K RMS of ECMWF's operational forecast at every lead to +168h, and 500 hPa geopotential within 3.6 m at +168h. At 8 US METAR stations it scored 1.68 °C RMSE against 1.66 °C for ECMWF's own run.

Attribution

AIFS Single 2.0 is © ECMWF, licensed under CC BY 4.0. The initial conditions are ECMWF IFS open data, also CC BY 4.0. The interpolation matrices are from ECMWF's earthkit-regrid, Apache 2.0. This wrapper's code is Apache 2.0.