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.
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.