graphcast_finetune_2019_2021
This model contains the GraphCast checkpoints created as part of (Subich 2024), which fine-tunes the "standard" GraphCast ¼°/37-level model on the 2019-2021 period. The primary goal of the study was to adapt the model to the Canadian GDPS analysis, but another product produced along the way was a "control" version trained on ERA5 data, which is more widely available.
The model's training code is available at https://github.com/csubich/graphcast.
The model checkpoints are in the params/ar{1,2,4,8,12} directories, each directory noting the number of autoregressive forecast steps completed. See the arxiv paper for details about the training schedule. The respective era5.ckpt files are the model versions trained on ERA5 data, and the gdps.ckpt files are those trained on the GDPS analysis data. The ar12 checkpoints are the final result of training, and the earlier ones are provided for research & reference.
The GDPS-tuned model was trained with an adjusted set of normalization weights, which are located in stats/gdps. For symmetry, the corresponding ERA5 weights are at stats/era5, but those are unmodified from the normalization weights used for the unmodified GraphCast models.
Also as noted in (Subich 2024), the models were trained with an alternate set of vertical (pressure level) weights for the loss function, which are included here in the various error_weights/*.pickle files. deepmind.pickle just reproduces pressure-proportional weighting, and it is included for completeness.
As these models are all derivative of the published 37-level GraphCast weights, these models also carry the CC-BY-NC-SA-4.0 (attribution, noncommercial, sharealike) license.