global-mini

The global-mini model is a small, global, 1-degree configuration designed to sit between a toy model and an operational-quality model.

  • It is not a toy. It uses the same pipeline, data tooling, and architecture family (Graph Transformer encoder/decoder with a Transformer processor) as the other EAGLE models.

  • It is not operational quality. Its reduced resolution, smaller network, and shorter training budget mean its skill is lower than the global-EAGLE and nested-EAGLE models.

It is trained so that inference can run on a CPU or a local laptop, with no GPU or HPC system required. This makes it useful for learning the workflow, testing code and configuration changes, training sessions, and demos.

Note

global-mini is intended for education and demonstration. For scientific or near-real-time forecasting, use the global-EAGLE or nested-EAGLE models.

Training Data

At a glance:

  • GFS initial conditions, conservatively regridded to 1 degree.

  • 6-hourly data.

  • The training period spans 2020-01-01 through 2024-12-31.

  • The validation period spans 2025-01-01 through 2025-12-31.

  • The testing period spans 2026-01-01 through 2026-09-30.

The model uses the same variable categories as the other EAGLE models:

Table 3 global-mini variables by category

Category

Fields

Diagnostic

u80, v80, accum_tp

Forcing

lsm, orog, cos_latitude, sin_latitude, cos_longitude, sin_longitude, cos_julian_day, sin_julian_day, cos_local_time, sin_local_time, insolation

All remaining dataset fields are treated as prognostic variables.

Model Architecture

Table 4 global-mini architecture

Component

Setting

Encoder and decoder

Graph Transformer (4 heads, MLP hidden ratio 2)

Processor

Transformer, 8 layers, 8 heads, window size 256, MLP hidden ratio 2

Channels

512

Latent mesh

Triangular mesh, refinement level 4 (2,562 nodes, roughly 250 km)

Attention backend

scaled_dot_product_attention (processor), pyg (encoder and decoder), so no GPU-specific kernels are needed

For comparison, nested-EAGLE uses a 16-layer processor with a window size of 4320 and requires GPU-specific kernels (flash attention and Triton).

Configuration

The model is configured in config/global_mini.yaml, which is composed on top of config/base.yaml:

uw config compose config/base.yaml config/global_mini.yaml > eagle.yaml

For Google Colab, an additional overlay, config/colab.yaml, reduces the time range and training length and removes the scheduler launcher so each step runs directly in the notebook:

uw config compose config/base.yaml config/global_mini.yaml config/colab.yaml > eagle.yaml

Demonstration Notebook

The notebooks/global_mini.ipynb notebook walks through the full EAGLE pipeline on a small dataset and can be run in Google Colab:

Open In Colab

The notebook covers:

  1. Environment setup: creates the base, data, anemoi, and wxvx conda environments.

  2. Config composition: builds the main config with uwtools.

  3. Data preprocessing: creates training and validation datasets from GFS with ufs2arco.

  4. Training: trains a graph-based model with anemoi-core modules.

  5. Inference: creates a forecast from the trained checkpoint with anemoi-inference and eagle-tools, then plots the output.

The notebook trains for only a few steps on a very short period, so the resulting forecasts are for demonstration of the workflow and are not meaningful in terms of skill. A T4 GPU is required for training in Colab.

Note

Planned notebook additions include loading a checkpoint pretrained for 40,000 steps, running near-real-time inference, and verifying the forecast with wxvx.