🌍 Weather Model Comparison

AIFS Single v2 (ECMWF) · WeatherNext 2 (Google DeepMind) · ERA5 climatology baseline

📋 Session Status

  • AIFS: ⬜ not run yet
  • WeatherNext2: ⬜ not run yet
  • Climatology (ERA5 baseline): ⬜ not run yet

⚙️ Run a Forecast

Model
AIFS runs on GPU (ZeroGPU); WeatherNext2 runs on CPU (~96s/step — needs ~50GB RAM/first-load, more than this Space's default GPU allocation); Climatology is a zero-skill baseline for comparison, not a real model.
Number of steps (6h each)
Initial conditions
Applies to whichever model you run — pick the same historical date for AIFS, WeatherNext2, and Climatology so their steps land on the same valid time and are actually comparable. Historical dates pull from ECMWF's deeper S3 archive (from 2023-01-18) instead of the live feed.
1 32

🗺️ Visualize

Model
Step
Field

📂 Saved Runs

Forecasts saved above persist in the weather-forecast-archive dataset — load one back here to plot/compare it without re-running the model.

Saved run

📊 Compare Two Models

Pick any two model runs (including two steps of the same model) and a field — shows both maps, their difference, and a skill metric (RMSE, MAE, bias, correlation). AIFS's irregular grid is compared by sampling the other model onto AIFS's own points; WeatherNext2 and the climatology baseline share an identical grid, so no resampling is needed between them.

Model A
Step A
Model B
Step B
Field

Notes

  • AIFS: no flash-attn required (PyTorch SDPA — works on CPU, MPS, CUDA). First run downloads its ~2GB checkpoint.
  • WeatherNext2: runs via an unmerged transformers fork; CPU-only here, ~96s/step once loaded.
  • Climatology baseline: pure ERA5 climatological mean per step — zero model skill by construction, a reference point for judging whether AIFS/WeatherNext2 add value.
  • Data: ECMWF Open Data (forecasts) and EarthMover's public ERA5 archive (climatology).