Model weights

Pretrained AuroraChem checkpoints and how to load them

Checkpoints

VersionTraining dataGridDownload
AuroraChem v1 GEOS-Chem (FETCH4) and ERA5, 2018 4°×5°, 13 pressure levels available on request
(posted upon publication)

AuroraChem v1 is the model described in Kim et al. (submitted). Each checkpoint has two files: the model weights (<run_name>_best_weights.pth) and the normalization statistics used during training (<run_name>_norm_stats.npz).

Loading a checkpoint

Place both files in output_dir/Weights/ and set run_name in configs/aurora_eval_config.yml to match the file names. The evaluation script loads the weights automatically:

python -m scripts.aurora_eval --config configs/aurora_eval_config.yml

See the Quick start for the full setup.

License

The AuroraChem code is released under the MIT License. The modified Aurora code in aurora/ is MIT-licensed by Microsoft.