Back to the lab

Open forecast studio

The model family

WeatherLab is an independent visualization of the WeatherNext-style ensemble suite — Functional Generative Networks, graph networks, diffusion, and the ECMWF physics baseline. Tracks and fields in the lab are a demonstration cycle initialized 00Z 20 August 2026. They are not official forecasts.

WeatherNext 2

WN2

Operational ensemble atmosphere. Faster than the previous generation, with tighter tails on extreme wind, rain, and cyclone structure.

Architecture
Functional Generative Network
Resolution
0.25° / 28 km
Ensemble
64–1,000 members
  • Eight times faster scenario generation than the prior family
  • Hourly-capable output; lab view steps at 6 hours
  • Joint track, intensity, and wind-structure skill on cyclones

WeatherNext Graph

Graph

Graph-network medium-range model. Strong on large-scale steering flow; slightly slower to deepen compact cores.

Architecture
Graph neural network
Resolution
0.25°
Ensemble
Probabilistic
  • Learns atmospheric state on a mesh rather than a regular grid
  • Often the left-of-track member in the Gulf this run
  • Useful as an independent steering check against WN2

WeatherNext Gen

Gen

Diffusion ensemble for alternate weather worlds. Wider spread, helpful when the synoptic pattern is unstable.

Architecture
Conditional diffusion
Resolution
0.25°
Ensemble
Diffusion ensemble
  • Samples the forecast distribution rather than a single trajectory
  • Tends to keep Hanna farther south in this cycle
  • Best read as a scenario generator, not a deterministic track

ECMWF ENS

ENS

The operational physics baseline. Shown so AI tracks can be read against the model centers already use.

Architecture
Physics IFS ensemble
Resolution
~18 km
Ensemble
51 members
  • Independent dynamical core — not an AI sibling
  • Favors earlier extratropical transition on Isaias
  • Official warnings still come from national centers, not this lab

Inspired by Google DeepMind WeatherNext and Weather Lab. This app does not run the official model weights. Point forecasts use Open-Meteo. Cyclone ensembles are a deterministic demonstration generated from a published-style 15-day envelope.