GraphCast loss function different from original DeepMind version

[cr-metdev]

I noticed that the graphcast loss function in this repository differs from what was used in the original graphcast paper, since the implemented here does not weight loss based on pressure level or type of variable. Just wondering if there was a reason for this or if there are plans to implement the level/variable weighting in the graphcast loss function in future updates?

[mnabian]

Hi @cr-metdev , thanks for posting your question. We started with a simplified version of the loss function, but now in a recent PR, the implementation has improved to include other loss weightings based on pressure level or type of variable: https://github.com/NVIDIA/modulus/pull/554

[mnabian]

Closing as complete