Possibility to have solar and long wave radiation output

[uturuncoglu]

I am trying to use GraphCastSmall model to do an experiment with WB2ERA5 data source and I wonder if there is a possibility to include extra output to the model run such as shortwave (net would be sufficient) and long wave radiation (downward component)? If it is not possible, is there any other model that could output those variables.

[NickGeneva]

Hey @uturuncoglu

The model outputs are fixed based on what the checkpoint provides, theres no way to add or remove input/output fields at will.
You can look at taking the output of GC and then plugging that into a diagnostic model that then predicts the radiation fields you are interested in.

Current supported diagnostic models are listed out here:
https://nvidia.github.io/earth2studio/modules/models.html#earth2studio-models-dx-diagnostic

Theres a few solar radiation ones listed
https://nvidia.github.io/earth2studio/modules/generated/models/dx/earth2studio.models.dx.SolarRadiationAFNO1H.html
https://nvidia.github.io/earth2studio/modules/generated/models/dx/earth2studio.models.dx.SolarRadiationAFNO6H.html

You might also be interested in the cbottle infill model where you could provide the outputs of GC and let cbottle generatively predict the rest (includes some radiation quantiites)
https://nvidia.github.io/earth2studio/modules/generated/models/dx/earth2studio.models.dx.CBottleInfill.html

Cbottle variables here:
https://github.com/NVIDIA/earth2studio/blob/main/earth2studio/models/dx/cbottle_infill.py#L54
https://github.com/NVIDIA/earth2studio/blob/main/earth2studio/lexicon/cbottle.py

Example of cbottle infill:
https://nvidia.github.io/earth2studio/examples/15_cbottle_generation.html#sphx-glr-examples-15-cbottle-generation-py

Or you could train your own to map from some set of predicted fields from ERA5 that GC predicts to the quantities of interest but thats outside the scope of this package.

[uturuncoglu]

@NickGeneva Thanks for your kindly help and also information. If I understand correctly, to have surface radiation components I need to run set of models in a chained fashion. So, maybe like GC -> SolarRadiationAFNO1H -> CBottle. Right? I think that would work for me if I could able to access the required radiation components (the others are already provided by GC). At least, this would be a proof of concept for my application and if I need to have a custom model then I could try to customize GC using ERA5 dataset to have full set of required fields. I also noticed that GC outputs for every 6-hours. Is there any way to set the output interval. Probably, that is fixed and maybe I could try to use InterpModAFNO for temporal interpolation to access fields in desired time.

[NickGeneva]

If you are interested in cbottle you could run GC -> cbottle, you would provide the cBottle model the common variables that GC and Cbottle share (random side note, there are some limits with cbottle and its time range / need to SST)

Similar with SolarRadiationAFNO1H you would just run GC -> SolarRadiationAFNO1H (assuming all input variables needed by SolarRadiationAFNO1H are provided by GC which I havent checked myself)

No need to chain diagnostics probably in this case.

I also noticed that GC outputs for every 6-hours. Is there any way to set the output interval

You are correct it is fixed at 6 hours, and indeed InterpModAFNO could provide you hourly. Although the interpolation model was trained for use with SFNO. So just a heads up.

[uturuncoglu]

@NickGeneva Thanks for confirmation. That is really helpful. Let me try to experiment to what happens. So, If InterpModAFNO does not work, I could bring some custom temporal interpolation routine.

[uturuncoglu]

@NickGeneva I wonder if there is an example to pass GC output to bottle. At this point, I am looking into https://nvidia.github.io/earth2studio/modules/generated/models/dx/earth2studio.models.dx.CBottleInfill.html but could not see any actual input rather than SST. It gets lit of input variables but I think it will download somehow but in my case I need to provide GC output. any idea?

[NickGeneva]

Right, thats the requirement with CBottle is that it needs SST and was trained on a limited past range.
There are two options:

1. Run cbottle for historical times (before 2022) where it will auto load the sst data that it was used to train with for you (no SST input needed). Allowing you to infile with any of the variables. In this case it uses a AMIP SST data that is monthly sst it will interpolate between internally.
2. Add your own SST source either via some forcing or model.

[NickGeneva]

Relevant code: https://github.com/NVIDIA/earth2studio/blob/main/earth2studio/models/dx/cbottle_infill.py#L433