[Paper] GraphCast: Learning skillful medium-range global weather forecasting

[jacobbieker]

Detailed Description

https://arxiv.org/abs/2212.12794

Context

Possible Implementation

[jacobbieker]

This paper is similar to Kiesler, but has a few changes, including:

1. Taking the previous timestep as well, so two timestep input, to output the next one

2. Multiple meshes for the latent graph, from super coarse (12 nodes and 20 faces) down to 40,962 nodes and 81,920 faces on highest resolution (6 total). Each scale is connected to the one above it, as each node is a subnode of a lower resolution one. Edges from all levels of the hierarchy are mapped onto the finest-resolution mesh
   

3. Data is stored in pressure levels as well, so is a 3D graph

4. Processor is 16-layer deep GNN

5. Loss is MSE with some modifications
   

[jacobbieker]

There seems like there might be an implementation here: https://github.com/HFAiLab/OpenCastKit

[mnabian]

We also have an implementation of GraphCast in NVIDIA Modulus.
Model: https://github.com/NVIDIA/modulus/tree/main/modulus/models/graphcast
Training recipe: https://github.com/NVIDIA/modulus-launch/tree/main/examples/weather/graphcast
Let me know if you had any questions!

[jacobbieker]

Thanks @mnabian! That looks really cool, I'll definitely check it out

[mnabian]

FYI, DeepMind released the GraphCast code and checkpoints a couple of days ago: https://github.com/deepmind/graphcast

[jacobbieker]

Awesome! Glad they released it.