AIFS Model

[jacobbieker]

https://www.ecmwf.int/en/about/media-centre/aifs-blog/2024/first-update-aifs

Detailed Description

AIFS model code isn't public, so not exactly sure how it is working, although its apparently similar to GraphCast. On a new update, it uses a reduced Gaussian octahedral grid for the processor, and the encoder an decoder use attention-based graph neural networks. The processor grid has 40320 grid points, which is processes as a sequence with a sliding attention window.

Context

It is another approach that might be worth including here. Good use for #76, as it could be another internal graph representation to easily slot in and try out.

Possible Implementation