Original title: AICON: An operational global machine learning weather forecasting model
Tobias Goecke, Marek Jacob, Florian Prill, Michael Denhard +8 · August 2026
Weather models often blur out tiny details when trained to forecast many days ahead. To solve this, researchers built AICON, a global machine learning weather forecasting model. Deutscher Wetterdienst put it into full operational use on 2nd of March 2026. AICON runs on a graph neural networkA type of model that works on a network of connected points, where each point's value is updated using its neighbors., which treats global weather data like a giant connected web. It uses an encoder-processor-decoder setup that updates connections with an attention mechanism. To match its training data, it maps the globe using an icosahedral multi-mesh grid and a special vertical coordinate that follows the terrain. During training, the team avoided multi-step rollouts and longer forecast horizons to keep fine-scale features sharp. Evaluations, including a tropical cyclone case, spectral analysis, and checks against real-world observations, showed its strengths and limits across different scales. Crucially, it matches the skill of the operational ICON model, especially for near-surface variables in short to medium ranges.
Imagine learning to paint a landscape. If you only practice the whole scene in one go, small details like leaves or ripples get lost; you get an impression but not the fine bits. So you might practice small sections in close-up, one at a time. In this paper, the weather model trains the same way: short local steps rather than long multi-step forecasts, so small-scale features survive.
National weather forecasts guide flight planning, ship routing, and severe-weather warnings. This model runs in real operations at Germany's Deutscher Wetterdienst, predicting at 13 km resolution with a 3-hour step, and is competitive with the traditional physics-based ICON model for near-surface conditions in the short-to-medium range. That makes data-driven forecasting a live option for national services, not just a research curiosity.
The abstract doesn't say what the specific weaknesses are. It only notes that evaluation reveals strengths and limitations in representing atmospheric variability across scales, and that skill is particularly strong for near-surface variables in the short to medium range—implying other variables or longer horizons may be less competitive, but that is not explicitly stated.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
AICON: An operational global machine learning weather forecasting model — Tobias Goecke, Marek Jacob, Florian Prill et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.