create GraphCast as a service

[liunelson]

The goal is to investigate how to implement the GraphCast ML model as service:

• Create a new repo containing the solution
• in a Docker container
• on demand
• with GPU enabled if available
• with job queue (since each run can be minutes to hours)
• eventually generalizable to the other 3 weather-forecasting ML models from the European Centre for Weather Forecasting (panguweather, fourcastnet, fourcastnetv2).

GraphCast can be installed in two ways:

1. via the ECWF's ai-models package https://github.com/ecmwf-lab/ai-models
2. via the official Google DeepMind graphcast package (https://github.com/google-deepmind/graphcast)[https://github.com/google-deepmind/graphcast]

I suggest (2) initially.

A working instance of GraphCast can be found here for reference. Note that you'll need to use a personal Google account to actually execute this notebook due to some Uncharted IT limitation.

This service would basically do as in the above Google Colab instance:

1. user specifies one of three "checkpoints" (GraphCast, GraphCast_small, GraphCast_operational)
2. Load this model into memory
3. user provides the input/initial dataset from which forecasting is made; currently, it is a selection from a list of available datasets; eventually, it should be a URL pointing to some dataset in a DB
4. Load this dataset into memory
5. Run this model with this dataset for N steps.

[liunelson]

@bigglesandginger Does this make sense to you?