3. Generate a Forecast

3.1. anemoi-inference Overview

We use the anemoi inference package to create a forecast. The anemoi-inference package helps users take trained ML models and generate forecasts for production or evaluation settings. Overall, it provides the workflows required to take trained ML models into real-world use.

See anemoi-inference documentation for further information.

3.2. anemoi-inference Quick Tips

anemoi-inference requires a YAML configuration to run via the CLI. A simplified YAML configuration looks like:

checkpoint: path/to/inference-last.cpkt
lead_time: 240 # hours
date: 2026-01-01T00

# lots of input options (see anemoi-inference documentation)
input: my_hrrr_initial_conditions.zarr

# lots of output options (see anemoi-inference documentation)
output: 2026-01-01T00.240hr.nc

This simple setup will successfully execute inference.

3.3. Using the Official Nested-EAGLE Checkpoint

The official pretrained nested-EAGLE checkpoint can be downloaded with:

wget -O inference-last.ckpt https://eaglecheckpoints.blob.core.windows.net/eagle-checkpoints/nested-eagle/inference-last.ckpt

To use this checkpoint for inference, set inference.anemoi.checkpoint_path to the downloaded file:

inference:
  anemoi:
    checkpoint_path: /path/to/inference-last.ckpt

If inference.checkpoint_dir is also present, the Inference driver will use the latest checkpoint from that directory instead of the explicit checkpoint_path. Remove or override checkpoint_dir when you want to force use of a specific pretrained checkpoint.