Models and checkpoints¶
The companion model repository is simon-donike/geo2wf-models on Hugging Face, paired with the scientific dataset. The paper compares field reconstruction, post-hoc scalar correction, and joint field/scalar learning.
| Model | Inputs | Outputs |
|---|---|---|
| Field U-Net | GEO, deterministic context, optional ERA5 | Surface wind-speed field; intensity/radii diagnosed from the image |
| Joint latent MLP | Same observation inputs | Field plus directly predicted maximum wind and optional radii |
| Encoder-only MLP | Same observation inputs | Maximum wind and optional radii, without a field decoder or SAR loss |
| Post-hoc correction | Frozen U-Net field and current metadata | Corrected maximum wind, optionally scalar radii |
| Six-hour forecast | Current intensity and recent intensity history | +6 h maximum wind; recursive +12 h diagnostic |
SAR is a training target for field models. Best-track intensity and radii supervise scalar heads. The forecast is a separate downstream task from the paper's instantaneous reconstruction comparisons.
Checkpoints¶
The 2 October 2026 release contains 21 PyTorch Lightning checkpoints (1.27 GB of checkpoint files), original configurations, provenance, and a matching source archive. Checkpoint binaries are not bundled in Git.
From the repository root:
uv tool install huggingface_hub
python3 scripts/download_artifacts.py models --dry-run
python3 scripts/download_artifacts.py models
Files are saved to downloads/models/. The script pins a published model commit and the dataset commit referenced by its dataset-links.json. Use all instead of models to fetch both releases. See the download guide for storage, selective data downloads, and using the matching source.
| Release path | Contents |
|---|---|
checkpoints/ | Released model weights and pretrained initializers |
release/registry.json | Checkpoint paths, hashes, original configurations, and experiment relationships |
dataset-links.json | Matching data revision, cohorts, normalization, and split hashes |
run-provenance/ | Original training and source evidence |
code/conference-source.tar.gz | Matching implementation, catalog loaders, reproduction scripts, and dependency lock |
The model card lists the checkpoint groups and validation limits. External dashboard ViT and ConvLSTM weights are not included; their available exported results are in the dataset.
Use each checkpoint with its matching resolved configuration, channel order, normalization statistics, and data cohort. Current presets are starting points for new runs and need not match a historical checkpoint. Record the full Hub commit for both data and models when reproducing results.
The installed inference and evaluation commands operate on local files. Full-state resume and weights-only initialization are covered in training.
Training presets¶
Names below are files under configs/experiment/, passed as experiment=<name> to geo2wf-train.
| Prediction | With ERA5 | Without ERA5 |
|---|---|---|
| Field only | intensity_comparison_unet | intensity_comparison_unet_no_era5 |
| Joint field and intensity | bottleneck_unet_mlp | bottleneck_unet_mlp_no_era5 |
| Joint field, intensity, and radii | latent_mlp_sar_era5_max_wind_radii | latent_mlp_sar_no_era5_max_wind_radii |
| Scalar intensity and radii, no SAR loss | latent_mlp_no_sar_era5_max_wind_radii | latent_mlp_no_sar_no_era5_max_wind_radii |
The latent_mlp_*_max_wind counterparts disable the radius head. Post-hoc correction uses unet_intensity_correction, with optional _image_radii and _mlp_radii variants. Forecast pretraining uses intensity_forecast_pretrain.
The no-ERA5 comparison presets keep ERA5 availability filtering to preserve the matched cohort while withholding ERA5 from the model. For a new study using observations without ERA5, set data.require_era5=false and document the changed cohort.