Question on extracting Aurora latent states

[moyan-liu]

Hi Aurora team,

I am currently working on using aurora for weather prediction, and thank you very much for the impressive work you’ve done on the model.

We are now exploring a way on how to extract Aurora's latent state at each time step and then to have a better understanding of the predictability. For example, which latent states are most important at different times or flow regimes, and how the relative importance of these latent states evolves over time.

Is this feasible and if so is it possible to recommend some ways to obtain these latent states from Aurora?

Thank you again for your work and your time.

[wesselb]

Hey @moyan-liu! Thanks for opening an issue. I'm glad you find the model useful. :)

The architecture of Aurora can be divided into three parts: an encoder, which compresses the input data into a latent representation; the backbone, which processes the latent representation; and a decoder, which converts the output of the backbone into the final forecasts.

For your purposes, it might be interesting to inspect the latent representation after the encoder, and possibly also after the backbone. The output of the encoder can be obtained here.