Adding Aurora to WeatherBench

[raspstephan]

Hi,

great work on the paper and also publishing the inference code. As you might know I am running the WeatherBench benchmark: https://sites.research.google/weatherbench/

I am currently in the process of updating WeatherBench with additional models and also updating to 2022. I think it would be great to have Aurora on there as well. If I had unlimited time, I would simply run your inference pipeline and save the data but unfortunately this isn't the case. So I wanted to ask whether there is appetite from your side to work together. Specifically, what I would need is the model output stored for 2022 (since I believe 2020 is part of your training?) for all 00/12 initializations from ERA5. The format isn't that important. Converting to the WB-standard Zarr format is something we can take on.

It would be awesome to also have the 0.1 deg HRES version but I believe that is trained up to incl. 2022 and, for the moment, 2022 is the most recent year we are using in WB.

Please let me know if you have any questions.

[wesselb]

Hey @raspstephan! Thanks for opening an issue. :)

I think it would be really cool to add Aurora as another baseline in WeatherBench! We might be able to send you predictions for 2022. Let me look into this.

Would using HRES T0 as initial condition instead of ERA5 be OK too? We don't have a model fine-tuned to ERA5 specifically. Our best model at 0.25 resolution is fine-tuned for HRES T0.

You're right that the 0.1 model is trained on 2022, so unfortunately that wouldn't work.

[raspstephan]

HRES t0 inits are perfect actually.

[wesselb]

@raspstephan I will reach out over email to continue the discussion there.