Will the presence of anomalous values cause errors in GraphCast's predictions?

[chufall]

Hi
As the comment in the GC code said : "The inner predictor is given inputs that are normalized using locations
and scales to roughly zero-mean unit variance."
If there're any anomalous values , will it cause errors in GraphCast's predictions ？
If it's so , how to fix it?
Thanks a lot
Sincerely.
Qc

[alvarosg]

That is generally an empirical question. If the anomalous values fall outside the distribution of training data (e.g. historical weather) then the model should in principle work as expected. Of course the training distribution may not be gaussian, and have longer tails, and it should still be fine to get some points that are in the tails (top bottom percentiles) of the distribution, and the model should do something reasonable so long as the overall state seems to also be within distribution. But if you do something like take a weather state, are 2 degrees to the temperature everywhere and try to run the model on that, then you would be clearly out of distribution state, and it is hard to predict what the model would output.