search

google-research / neuralgcm

Hybrid ML + physics model of the Earth's atmosphere
https://neuralgcm.readthedocs.io
Apache License 2.0
698 stars 79 forks source link

Forcing NeuralGCM with monthly mean data #137

Open smhenry2 opened 3 weeks ago

smhenry2 commented 3 weeks ago

From the NeuralGCM documentation, "all_forcings" uses the function "forcings_from_xarray" for all ERA5 data to create the forcings. It is stated that “ERA5’s sea surface temperature and sea ice concentration variables exclude the diurnal cycle and are updated only about once daily, so passing in hourly forcings is unnecessary”. I am interested to know 1) how the SST and sea ice data are updated by NeuralGCM when using a theoretical monthly forcing data, and 2) how to pass monthly mean data as forcings given that the number of days varies between months. Could the authors provide more information? Thanks.

yaniyuval commented 2 weeks ago

Hi @smhenry2 , I am not sure I fully understand your questions.

smhenry2 commented 2 weeks ago

Hi @smhenry2 , I am not sure I fully understand your questions.

Hello @yaniyuval, To clarify, I would like to know how to best force NeuralGCM with forcing data of a different time resolution than model output time resolution for a dynamic forcing. For instance, in the demo Jupyter Notebook, both the ERA5 forcing data and the prediction output are of time resolution 24h. How could dynamic forcing be used to force NeuralGCM with weekly or monthly data while maintaining 24 hourly output data, and how might that impact the predictions? Sorry if my initial question was not clear enough.

shoyer commented 1 day ago

We haven't experimented much with this. In principle it should work, but I don't know the best settings to use.