WEATHERMAPS.AI · THE AI MODEL DESK OF THE FORECAST NETWORK FORECAST EAGLE FORECAST EURO WEATHER.CITY WEATHER.FOOTBALL
weathermaps.ai
The Board · every model, every run, on one page

The atmosphere, learned and solved, side by side

weathermaps.ai tracks the machine-learning weather models against the physics models they are measured against. One model's map carries almost no information, because every model draws a plausible Tuesday. What carries information is where two of them, run at the same hour, stop agreeing.

Desk Note · 31 Oct 12Z run

Novosibirsk tops a run where model disagreement grows with lead time

In the 2026-10-31 12Z run, 602 cities were scored. The 90th percentile of city scores rises from 30 on day 1 to 46 on day 3, then reaches 53 on day 7 and day 10. The median city scores 36 on day 5 and 41 on day 10. The most contested city is Novosibirsk, Russia, on day 5, with a DI of 70 or more: 74, in the high band. Temperature drives it. Day highs are 2 °C for AIFS, 2 °C for HRES, 1 °C for WeatherNext 2 and 2 °C for the ensemble mean.

Lyon, France, scores 70 on day 5, driven by temperature. AIFS has 8 °C, HRES 6 °C, WeatherNext 2 8 °C and the ensemble mean 9 °C. Edmonton, Canada, scores 68 on day 5, also driven by temperature: AIFS shows 8 °C, against 4 °C for HRES and 3 °C for both WeatherNext 2 and the ensemble mean. On day 7, pressure drives the top scores in Taipei (65), Daejeon (64) and Seoul (64).

Every feed reports ok and a 12Z newest run except IFS ENS, which is late and shows 00Z as its newest run. GFS shows 18Z. GraphCast, Aurora, FourCastNet v2 and Pangu-Weather, run by NOAA from the ECMWF analysis, also have GFS-start runs at 12Z.

Machine-written by Claude from these numbers, and published only after every number and city in it was checked against them. It describes where the models disagree; it is not a forecast.

On the Board 9 of 14 tracked 6 learned · 3 physics
Newest cycle 31 Oct 18Z 10 h ago
Feeds on time 9 / 9 Status page →
Disagreement · day 5 50 NOTABLE 90th percentile of 602 cities · 31 Oct 12Z run

Most contested cities

Where the 8 models disagree most, measured against how much the ensemble itself expects. From the 31 Oct 12Z run. All 602 cities →

50 = as much as the ensemble expects
City DI Driven by Highs, °C to Thu 5 12Z
NovosibirskRussia 74 temperature 2212 spread 1°
LyonFrance 70 temperature 8689 spread 3°
EdmontonCanada 68 temperature 8433 spread 5°
BernSwitzerland 64 temperature 7768 spread 2°
CalgaryCanada 64 temperature 8565 spread 3°
OttawaCanada 64 temperature 20181917 spread 3°
OsloNorway 63 temperature 5657 spread 2°
YerevanArmenia 62 pressure 11131414 spread 3°
Las VegasUnited States 61 temperature 19181817 spread 2°
GenevaSwitzerland 59 temperature 9989 spread 1°
MontréalCanada 59 temperature 19161917 spread 3°
AshgabatTurkmenistan 57 pressure 23242423 spread 1°

Runs coming in

Each model's newest run and how far through its forecast the rendered maps have reached. Unfinished runs come first.

Model Board

Click a column to sort it. ECMWF, DeepMind and NOAA models open in the explorer; the four NOAA-run learned models have city data and scores but no maps yet.

updated 1 Nov 04:03Z
Model Status Newest run Horizon Fields Grid Rendered by
AIFS SingleECMWF · Deterministic On time 31 Oct 18Z10 h ago f360 11 0.25° forecasteuro.com
WeatherNext 2Google DeepMind · Ensemble (64 members) On timeExperimental 31 Oct 18Z10 h ago f360 4 0.25° forecasteuro.com
GraphCastGoogle DeepMind · run by NOAA · Deterministic On timeGFS start: 31 Oct 12Z 31 Oct 12Z16 h ago f240 2 0.25° NOAA OAR
AuroraMicrosoft Research · run by NOAA · Foundation model On timeGFS start: 31 Oct 12Z 31 Oct 12Z16 h ago f240 2 0.25° NOAA OAR
FourCastNet v2 (SFNO)NVIDIA · run by NOAA · Deterministic On timeGFS start: 31 Oct 12Z 31 Oct 12Z16 h ago f240 2 0.25° NOAA OAR
Pangu-WeatherHuawei · run by NOAA · Deterministic On timeGFS start: 31 Oct 12Z 31 Oct 12Z16 h ago f240 2 0.25° NOAA OAR
IFS HRESECMWF · Deterministic On time 31 Oct 18Z10 h ago f144 23 0.1° forecasteuro.com
IFS ENSECMWF · Ensemble On time 31 Oct 18Z10 h ago f144 72 0.25° forecasteuro.com
GFSNOAA NCEP · Deterministic On time 31 Oct 18Z10 h ago f240 8 0.117° forecasteagle.com

Open a model

Every explorer uses the same colours, places and timeline, so switching models changes only the forecast.

Learned

AIFS Single ECMWF WeatherNext 2 Google DeepMind · experimental GraphCast Google DeepMind Aurora Microsoft Research FourCastNet v2 (SFNO) NVIDIA Pangu-Weather Huawei Every learned model we track 11 on the roster

Physics

IFS HRES ECMWF IFS ENS ECMWF GFS NOAA NCEP · on forecasteagle.com

Compare

AIFS against HRES Map options → Compare & verify WeatherNext 2 against HRES same valid time Where the ensemble is unsure 2 m temperature spread

Why this site exists

The baseline moved

For fifty years, better forecasts meant a better dynamical core and a bigger machine. Then a model that had never been told what a pressure gradient is started beating one built out of them. That is worth watching closely rather than through a press release.

Difference is the signal

One model's map tells you almost nothing. What carries information is where two of them, run at the same hour, stop agreeing, so this site is built around comparing models rather than around any one of them.

Coarse, and honest about it

The global learned models here run at a quarter degree, about 28 km. Where detail appears at a finer scale than the grid, it was generated rather than forecast, and the page says which.