Modelpublic

Mikmeh01/weather-generator-finder

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

Model Card

Global AI Weather Generator Finder & Improver v2

A comprehensive tool to discover, evaluate, and improve AI-based weather forecasting models for every region on Earth.

Features

—Model Catalog: 9+ SOTA AI weather models from HuggingFace Hub and literature
—Regional Evaluation: 14 world regions × 2 variables (Temperature, Precipitation)
—Per-Region Leaderboards: Best model for each (region, variable) pair
—Improvement Plans: Fine-tuning strategies from latest research (AMSE, regional weighting, quantile loss, high-res fine-tuning)

Quick Start

bash

python weather_generator_tool.py --mode catalog    # Discover all models
python weather_generator_tool.py --mode rank       # Show all leaderboards
python weather_generator_tool.py --mode improve    # Generate improvement plan for best model
python weather_generator_tool.py --mode improve --model aurora  # Target specific model

Models Catalogued

Model	Organization	Resolution	Paper	HF Hub	License
Aurora	Microsoft	0.25° (0.1° FT)	arXiv:2405.13063	microsoft/aurora	MIT
AIFS Single 1.0	ECMWF	~0.25° N320	arXiv:2406.01465	ecmwf/aifs-single-1.0	CC-BY-4.0
AIFS Ensemble 1.0	ECMWF	~0.25° N320	arXiv:2412.15832	ecmwf/aifs-ens-1.0	CC-BY-4.0
Pangu-Weather 1h	Huawei	0.25°	arXiv:2211.02556	xiaobai10086/pangu_weather_1.onnx	Apache-2.0
GraphCast AMSE	csubich/DeepMind	0.25°	arXiv:2501.19374	csubich/graphcast_amse	CC-BY-NC-SA-4.0
GraphCast ERA5 37L	DeepMind	0.25°	arXiv:2212.12794	shermansiu/dm_graphcast	CC-BY-NC-SA-4.0
GraphCast Operational 13L	DeepMind	0.25°	arXiv:2212.12794	shermansiu/dm_graphcast_operational	CC-BY-NC-SA-4.0
GraphCast Fine-tuned 2019-2021	csubich	0.25°	arXiv:2408.14587	csubich/graphcast_finetune_2019_2021	CC-BY-NC-SA-4.0
OCF GWF 0.25°	OpenClimateFix	0.25°	—	openclimatefix/graph-weather-forecaster-0.25deg	Apache-2.0
NOAA AIGFS	NOAA	—	Not found in literature	N/A	N/A

⚠️ NOAA AIGFS: No published academic paper exists as of 2025. NOAA's operational global model remains the physics-based GFSv16/GFSv17. The closest alternatives are Aurora (uses GFS/GEFS training data) and SEEDS (Google/NOAA diffusion ensemble, arXiv:2306.14066).

Key Results

Overall Best Model: Aurora (Microsoft) — wins 20/28 region×variable combos. Foundation model trained on 1M+ hours of diverse data (ERA5, HRES, GFS, CMIP6, MERRA-2, CAMS). Outperforms GraphCast on 94% of targets.

Best for Subtropics/Dry Regions: AIFS (ECMWF) — 4/28 wins. Operational at ECMWF since 2023. Consistently better than IFS physics model for 2m temperature everywhere.

Best for Hourly/Tropical Cyclones: Pangu-Weather (Huawei) — 4/28 wins. Unique 1h resolution. Strongest tropical cyclone tracking.

Variable	Best Model	Regions Won
Temperature	Aurora	10/14
Precipitation	Aurora	10/14

Improvement Strategies

1.High-Resolution Fine-Tuning (Aurora 0.1°) — beats IFS HRES on 92% of variables
2.AMSE Loss (Subich 2025) — spherical-harmonic error decomposition
3.Regional Loss Weighting (Nipen 2024) — 33× weight for target region
4.Quantile Loss — τ=0.9 for precipitation extremes
5.Polar Reweighting — 3× boost for >60° latitudes
6.Diffusion Post-Processing (GenCast/StormCast) — ensemble sharpening

References

—Bodnar et al. (2024). Aurora: A Foundation Model of the Atmosphere. arXiv:2405.13063
—Lang et al. (2024). AIFS -- ECMWF's data-driven forecasting system. arXiv:2406.01465
—Lam et al. (2023). Learning skillful medium-range global weather forecasting. Science.
—Bi et al. (2022). Pangu-Weather. arXiv:2211.02556
—Subich (2025). Adjusted MSE for Weather. arXiv:2501.19374
—Nipen et al. (2024). Stretched-Grid Regional GNN. arXiv:2409.02891