A.I. Is Quietly Powering a Revolution in Weather Prediction (e360.yale.edu)

In February the European Centre for Medium‑Range Weather Forecasts (ECMWF) quietly put into operation the world’s first fully operational, AI‑powered global forecast system (AIFS). The move marks a practical shift: machine‑learning forecasts can be produced far faster, cheaper and with roughly 1,000× less computational energy than traditional numerical models, while matching or exceeding accuracy in many cases (ECMWF reports ~20% better scores for some phenomena). Parallel research systems—GraphCast, FourCast, Pangu‑Weather and Google’s new ensemble GenCast—have shown comparable or superior skill to classical models and, in GenCast’s case, deliver probabilistic ensembles (50+ members) that outcompete existing physics‑based ensembles. Technically, most AI forecasters are trained on ERA5 reanalysis data and currently rely on outputs from numerical models to generate training targets, so physics‑based systems remain essential input sources. AI forecasts tend to be coarser (ECMWF’s are ~3× coarser) and can struggle with small‑scale surface winds and out‑of‑sample extremes, though tests (e.g., Storm Ciarán, Oct 2023) were promising. Key implications for ML: massive computational efficiency and democratization of forecasting for lower‑resourced regions; urgent needs for robust observational data pipelines, explainable AI methods, ensemble approaches, and continued evaluation under climate‑shifted conditions before AI can fully supplant physical models.