Image: James Hutchinson ( CC BY-SA 4.0 )
What it is
ECMWF’s Artificial Intelligence Forecasting System (AIFS) is a machine-learned global forecast that runs in production next to the physics-based IFS. AIFS Single 1.1.0 went operational on 27 August 2025. A 2026 Geoscientific Model Development paper describes a bounding layer that enforces physical limits (no negative rain, consistent cloud and precipitation fields) and reports about 4-6 percent better upper-air and near-surface skill, with the largest precipitation gains: up to 12 percent and about a day’s advantage on a categorical score. ECMWF later superseded 1.1 with AIFS v2 in May 2026; this card keeps the 1.1 line the founder asked for, labeled as that vintage. AIFS is operational software, not a demo.
Problem it targets
Physics models are accurate and expensive. A fast learned model can fill the gap if it stays physically sane and is scored against the same observations.
How it works
A graph-neural-network encoder, transformer processor, and decoder, trained on ERA5 and ECMWF analyses, steps the atmosphere in six-hour leaps. Four runs a day. Ten-day fields in minutes on a GPU is the operational selling point; the VAST Data essay the founder cited is a secondary telling of that shift. The legal and scientific record is the GMD paper and ECMWF’s version history.
Status and players
ECMWF; EuroHPC training machines (Leonardo, MareNostrum 5). Neighbor card: GraphCast and WeatherNext 3. IFS ensembles still handle spread and long coupling.
Risks and limits
Learned models can invent polite extremes or break conservation if unconstrained: that is why 1.1 added bounds. GPU non-determinism means you cannot bit-reproduce an official run at home. AI weather still needs the satellites.
Sources
Founder link: VAST Data essay on AIFS 1.1. Skill numbers follow Moldovan et al., GMD, 2026. ECMWF superseded 1.1 with v2 in May 2026.