Graphcast Operational training protocol

[tobiasselz]

Hello!

As I understand, the operational version of GraphCast is pretrained from scratch on only 13 levels on ERA5 data and then fine-tuned on IFS-HRES data. But other than this, is the training protocol similar to the one described in the paper? In particular, is the pre-training and/or fine-tuning also done autoregressively over 12 time steps (3 days)?

Thanks a lot and kind regards,
Tobias

[andrewlkd]

Hi Tobias,

It is trained with:

• 300k steps @ 1AR on ERA5
• 30k steps @ 1AR on HRES-fc0
• 1k steps @ 2-12AR each on HRES-fc0

Hope that helps,

Andrew