🌦️ AIFS Weather Forecasting Pipeline

Hugging Face model: ecmwf/aifs-single-1.0

This repository provides a Hugging Face–style inference and visualization pipeline for the ECMWF AIFS (Artificial Intelligence Forecasting System) weather model. It automates the process of:

• Fetching initialization data from ECMWF Open Data
• Preprocessing and regridding with earthkit
• Running AI forecasts using the Anemoi Inference framework
• Saving outputs to NetCDF files
• (Optional) Visualizing weather fields such as winds and temperature

🚀 Features

• ✅ Automatic initialization from ECMWF Open Data
• ✅ Compatible with ecmwf/aifs-single-1.0 checkpoints
• ✅ Multi-step AI forecast via Anemoi inference runners
• ✅ Output saving to .nc (NetCDF) format
• ✅ Optional map visualization with Cartopy

📦 Dependencies

pip install earthkit-data earthkit-regrid xarray matplotlib cartopy anemoi-inference ecmwf-opendata

⚙️ Usage

from aifs_pipeline import WeatherConfig, AIFSPipeline, main

# Initialize config (edit paths as needed)
config = WeatherConfig(
    checkpoint_path="/path/to/aifs-single-mse-1.0.ckpt",
    output_nc_base_path="./outputs"
)

# Run full pipeline
main(config)

This will:

1. Download the latest ECMWF open data
2. Initialize the model state
3. Run forecasts up to 360 hours ahead
4. Save results as .nc files under ./outputs

📊 Example Output

NetCDF files (e.g., 006.nc, 012.nc, ...) will be generated in:

outputs/

Each file contains model variables such as:

• 2t (2-meter temperature)
• 10u, 10v (10-meter winds)
• msl (mean sea-level pressure)
• z_500 (geopotential at 500 hPa), etc.

---

🖼️ Visualization Example

from aifs_pipeline import Visualizer, WeatherState

# Assuming 'state' is a WeatherState object
Visualizer(config).plot_state(state, save_path="wind_map.png", field_name="10u")

📚 Reference

• 🌐 ECMWF AIFS Model
• 📦 Anemoi Inference Library
• 🌍 ECMWF Open Data Access

📝 License

This project follows the same license as the ECMWF AIFS model.