WeatherNext Explorer

This is an experimental feed, not an operational product. It is a four-field subset of Google DeepMind's WeatherNext 2, mirrored here because the comparison it makes possible is worth having — but it carries none of the guarantees the ECMWF pages on this site do, and it should not be relied on the way an operational forecast can be. Everywhere it appears — the model chips, the compare menu, the polar explorer — it is marked.

WeatherNext 2 is a machine-learning ensemble, and what is drawn here is the mean of its members. That makes it the direct counterpart of two other things on this site at once: it is a data-driven model like AIFS, and it is an ensemble mean like the EPS — so it is smoothed in the same way and for the same reason, and past about day four that smoothing is a feature rather than a loss.

It is the first model on this site that is not ECMWF's, and that is the whole argument for carrying it. Every other comparison here is the Euro model against itself: HRES against its own AI counterpart, or either against its own ensemble. Those are real questions, but they share a centre, a reanalysis and a set of assumptions. A second centre's model shares none of that, so where it and ECMWF disagree the disagreement means something different — and quite possibly something more.

Four fields, and the pairings they were chosen for. 500 hPa height is the one to open first: it is the field nearly all medium-range skill is scored on, and it is the only field two ensembles on this site both publish, so Compare & verify against the ECMWF ensemble mean at day ten is the most interesting difference map here. 2-metre temperature, mean sea-level pressure and precipitation fill in the surface — and they are exactly the three the ECMWF ensemble page cannot show you, because ECMWF publishes no ensemble mean for them. So the two ensemble pages are complements rather than duplicates.

The grid is 0.25°, six-hourly to fifteen days, four cycles a day. Like the ECMWF ensemble it publishes no analysis frame, so the timeline starts at +6 h. Precipitation is an accumulation over each six-hour step — which is HRES's convention and the exact reverse of AIFS's, on another feed that also carries "machine learning" on the label; the value was measured off the data rather than taken from the variable name, as everything on this site is.

As on the other explorers, one raster covers the planet and every one of the 226 countries, twelve continental frames and eight ocean basins is a viewport on a field that is already drawn. The pointer reads a value back out of the picture and reports the ecCharts band it falls in rather than a precision a banded chart never claimed. The colours are ECMWF's own tables, used here deliberately on a model that is not ECMWF's: a difference between two forecasts only means something if nothing else on screen differs.

WeatherNext 2 © Google DeepMind. Forecast Euro is not affiliated with Google or with ECMWF.