[Paper] AI-based Data Assimilation

[jacobbieker]

Arxiv/Blog/Paper Link

https://arxiv.org/pdf/2406.00390

Detailed Description

Changes the data assimilation approach to be self-supervised and more similar to the physics assimilation process. The loss is the 3D-Var equation, and all that is needed is a first guess state and observations. The first guess state is generally a previous forecast step, but can also be blank if none exist.

They do this for 2m temperature as a real-world case over Germany and get interesting and promising results.

Context

Data assimilation is a important process, this could be an interesting way of doing it.

[aouatefm]

Hi @jacobbieker 👋
I’d love to contribute to this issue — the self-supervised data assimilation approach sounds fascinating.
Before I get started, could you please confirm whether the goal is to:

Reproduce the results from the paper (for 2m temperature over Germany), or

Integrate a similar AI-based assimilation method into the existing framework for experimentation?

Once I know the intended scope, I can outline an implementation plan or initial experiment setup and proceed with a PR.

[jacobbieker]

Hi!

Great that you are interested. The goal is to integrate a similar method into the existing framework. If you can reproduce the results, that would be great, but the main thing would be an implementation for this.

[SOHAMPAL23]

Hey @jacobbieker Can I work on this issue

[jacobbieker]

@SOHAMPAL23 sure!

[SACHokstack]

Hey @jacobbieker I would like to try this issue

[jacobbieker]

@SACHokstack I would maybe recommend a different issue, as @SOHAMPAL23 has just experessed interest. But if you want to continue anyway, feel free to!

[SOHAMPAL23]

@jacobbieker Working on the issue

[SOHAMPAL23]

#194

[jacobbieker]

#194

You should link to this issue from the PR description please.