How to Convert Aurora Model Output from [90→−90, 0→360] to Conventional [−90→90, −180→180] Coordinates?

[Mariia-Otset]

Hello @wesselb

I am currently forecasting 2-meter temperature using the Aurora model. The model outputs data on a global grid with the following configuration:
Latitude: ranges from 90° to −90° (descending order).
Longitude: ranges from 0° to 360°.
Data shape: [time, lat, lon] → [t, h, w].

For my application, I need to reformat the output to use the conventional geographic coordinate system:
Latitude: increasing from −90° to 90°.
Longitude: shifted to −180° to 180°.
Data shape remains [t, h, w].

What is the correct approach to transform the model output into this desired coordinate system?

[wesselb]

Hey @Mariia-Otset! Thanks for opening an issue. :)

Let x be a prediction for a surface-level variable, which is of the form (b, 1, h, w). To change the order of the latitudes, we flip along dimension -2 (the h dimension):

x = torch.flip(x, (-2,))
lat = torch.flip(prediction.metadata.lat, (0,))

Then, to rearrange the longitudes, first transforms the longitudes to the desired range and then sort according to those values:

lon = (prediction.metadata.lon + 180) % 360 - 180
inds = torch.argsort(lon)
x = x[..., inds]
lon = lon[inds]

I didn't test this myself, so you might have to debug it a little, but hopefully the idea is clear.

[Mariia-Otset]

Thank you so much @wesselb

[Mariia-Otset]

Hi @wesselb,
How can I use the aurora-0.25-pretrained model to forecast 2m dew point temperature?
Does it support predicting dew point temperature or do i have to fine-tuning this model?

[wesselb]

@Mariia-Otset, the pretrained version of Aurora 0.25 does not support 2 m dew point temperature. If you want to predict dew point temperatures, you will need to fine-tune the model to this variable. I suspect that this would work well!

[Mariia-Otset]

@wesselb
I’ve already fine-tuned and trained the model, but the loss is still quite high.
If I share the fine-tuning code, would you be able to review it and provide feedback?
Thank you in advance for your help.

[wesselb]

@Mariia-Otset I'm currently pretty busy and don't have the capacity to review full fine-tuning script, but I'm happy to answer specific questions or help with specific issues. :)