[Paper] CaFA: Global Weather Forecasting with Factorized Attention on Sphere

[jacobbieker]

Arxiv/Blog/Paper Link

https://arxiv.org/abs/2405.07395

Detailed Description

Context

The factorized attention is quadratic in axial attention rather than the full size, which should greatly reduce the computational power needed for the model, both training (more important) and inference, since most of the global weather models seem to be able to do inference on small amounts of compute, but training takes a lot of it.

[jacobbieker]

Has the same encoder-processor-decoder as many other models, so could fit well in this repo