The brief
- Runs AIGFS, an AI global model built by retraining DeepMind's GraphCast on NOAA's own weather analyses.
- A 16-day AIGFS forecast uses 0.3% of the computing of the physics GFS and finishes in about 40 minutes.
- HGEFS blends 31 AI and 31 physics forecasts; NOAA says no other forecaster runs such a hybrid ensemble.
Technical approach
Retrain an open AI forecaster
NOAA started from Google DeepMind's GraphCast and fine-tuned it on its own data-assimilation analyses, so it starts well from US initial conditions.
Mix AI and physics runs
The hybrid ensemble pools 31 AI members with 31 physics members; together they beat either set alone on most of NOAA's scores.
In-house AI program
The models came from Project EAGLE, a multi-year effort of the Weather Service, NOAA research labs, the modeling center and EPIC.
