AICON, a forecast learnt from ICON

Two routes that part tell you the forecast is uncertain, but not why. Models from different centres start from different pictures of the weather and compute it differently, and their gap mixes both. A pair that starts from the very same picture tells more: where its two routes part, the doubt lies in the forecast itself.

AICON, the machine-learned model of Germany's weather service, DWD, makes such a pair with ICON. It solves no equation: it learnt from fifteen years of ICON's own record of past weather how the atmosphere moves on, three hours at a time, and each of its runs starts from ICON's analysis. In service at DWD since March 2026, it now gives a route on every passage, beside ICON's.

It covers the whole world, a forecast every three hours up to seven and a half days ahead from its midnight and midday runs, five days from the others. Its strength is the wind near the surface in the first days: over a year of checks, DWD found its forecasts near the surface better than ICON's on most counts, and steadier from one run to the next, a forecast that flip-flops less while you wait for a window.

Say you plan to leave Kiel for Gotland in three days. If AICON's route and ICON's agree, the two models read the weather alike, though they may share the same starting error; if they part, the two models alone disagree, since both start from the same weather: a clean sign to wait for the next run before choosing. Only a model from another centre, GFS or ECMWF, can confirm their agreement: when it does, the window is as sure as a forecast gets.

Learnt to keep its average error small, AICON smooths: DWD's forecasters found that it underestimates strong winds, and in a typhoon its storm was far too weak. In a gale or along a sharp front, read its wind as too light and its timing as the useful part. Beyond about two days, its large weather systems deserve less trust than ICON's, and at a quarter of a degree it sees no sea breeze, no cape and no squall.